The World Bank Economic Review, 37(4), 2023, 519–548 https://doi.org10.1093/wber/lhad016 Article The Social Protection Engel Curve Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Michael Lokshin , Martin Ravallion, and Iván Torre Abstract Why do richer countries spend a higher share of their income on social protection than poor countries? A newly assembled dataset on social protection spending for 142 countries since 1995 allows an exploration of alternate hypotheses, treating the pandemic period separately, as it entailed a large expansion in social protection efforts. While the mean income share devoted to social protection rises with income, this is attributable to multiple confounders, including relative prices, weak governance in low-income countries, and access to information- communication technologies. Controlling for these, social protection spending is similar between rich and poor countries. This was also true during the pandemic. JEL classification: H53, I38, O15 Keywords: social protection, Engel curve, pandemic, governance, distribution, ICT, selective data reporting 1. Introduction Social protection (SP) has long been seen as an important task for governments in rich countries but much less so in poor ones.1 Public spending on SP tends to account for a higher share of national income in richer countries, implying an overall income elasticity exceeding unity.2 In short, SP spending appears to be a “luxury good” at the national level. Michael Lokshin (corresponding author) is Lead Economist at the Office of the Chief Economist for Europe and Central Asia at the World Bank; his email address is mlokshin@worldbank.org. Martin Ravallion was a professor at Georgetown University. Iván Torre is Senior Economist at the Office of the Chief Economist for Europe and Central Asia at the World Bank; his email address is itorre@worldbank.org. The research for this article was financed by the analytical program of the Office of the Chief Economist for Europe and Central Asia at the World Bank, and Georgetown University. The authors are grateful to Harold Alderman, Aart Kraay, Peter Lindert, Franco Peracchi, Nithin Umapathi, Frank Vella, Dominique van de Walle, and two anonymous referees for comments and/or discussions. Aylén Rodriguez Ferrari and Michael Gottschalk provided excellent assistance in data collection. A supplementary online appendix is available with this article at The World Bank Economic Review website. 1 Following World Bank usage, “SP” refers to public programs for “social assistance” (in the form of cash or in-kind benefits often targeted to poor families), “social insurance” (mainly unemployment benefits and pensions), and “active labor market programs” (such as job training schemes). On the history of thought on the role of SP, see Ravallion (2016a, Part 1). 2 See, e.g., Kristov, Lindert, and McClelland (1992); Lindert (1994); Peracchi (2001); Cornelisse and Goudswaard (2002); Auteri and Costantini (2004); Shelton (2007); Brückner, Chong, and Gradstein (2012); and Clemente, Marcuello, and Montañes (2012). The finding that public spending as a whole tends to rise as a share of GDP as GDP rises is sometimes referred to as Wagner’s Law. C 2023 International Bank for Reconstruction and Development / The World Bank. Published by Oxford University Press 520 Lokshin, Ravallion, and Torre Figure 1. Scatter Plot of Social Protection Spending as a Share of GDP against log GDP per Capita for the Latest Available Pre- Pandemic Year and the Nonparametric SPEC. 25 FRADNK FIN ITA SP expenditure, % of GDP (latest available year) AUT BEL Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 GRC SWENOR 20 DEU NLD CHE PRT SMR ESP ISL POL JPN LUX 15 UKR HRV GBR BRA BLR SRB CZE MDA LVA ISR CAN KGZ BIH ARG NZL 10 KWT EGY AUSUSA DZA CHN IRL MNG TUR YEM URY GEO SYC LSO UZB KOR TTO 5 MWI ZWE TLS AZE MEX MAC TJK SLV HKG ETH SEN KNA THA MOZ MAR BGD TUN ARE BDI AFG ATG BHR DOM BHS SLE MLI PHL JAM SGP COD LBR SDN 0 SOM 6.5 7 8 9 10 11 12 Log of GDP per capita (USD PPP) Source: Analysis based on social protection expenditure data is from the authors’ data set. GDP data from WDI. Note: The nonparametric regression line is a smoothed scatter plot (using the lowess command in Stata). The shaded area is the 95 percent confidence interval. The solid black line includes all the countries, and the dashed one excludes the five richest countries as measured by their log GDP per capita (Macao, Luxembourg, Singapore, Kuwait, and Ireland). The supplementary online appendix provides the corresponding graph deleting pension spending. Country codes are the U.N.’s Alpha 3 list.3 Figure 1 plots the relationship between the share of GDP devoted to SP and (log) GDP per capita across countries, mostly for 2019.4 The line of best fit is an example of what can be called the Social Protection Engel Curve (SPEC for short), defined as the mean share of national income devoted to SP, conditional on national income.5 The SPEC rises through almost the whole range of GDP (only turning down when reaching the five countries with the highest GDP per capita). A large difference is seen between rich and poor countries in the mean shares of GDP devoted to SP; the share is 15 percent on average among the top quintile of countries ranked by GDP per capita (albeit with a large variance), as compared to around 3 percent for the bottom quintile. 3 “Country Codes List,” https://www.nationsonline.org/oneworld/country_code_list.htm#R. 4 Figure 1 uses the most recent pre-pandemic data on total SP spending for each country. The analysis considers the pandemic period separately. It also considers the effect of dropping spending on pensions; the pattern in fig. 1 is very similar in this case (see supplementary online appendix). 5 As estimated by a nonparametric regression function in fig. 1. The study’s use of the term “Engel curve” borrows from the more familiar usage in the context of modeling consumer demands. This is not the first use of the term in the context of public spending, which appears to be Bird (1971). Figure 1 and other SP Engel curves in this paper are nonlinear (and non-parametric) generalizations of the longstanding Working-Leser specification for such Engel curves, for which the share of total income (or total spending) is taken to be linear in log total income (Working 1943; Leser 1963). The World Bank Economic Review 521 Evidence such as fig. 1 has been used to argue that poor countries have higher priorities than SP. Indeed, SP is not even mentioned in the classic policy package for developing countries that emerged in the 1990s, often referred to as the “Washington Consensus,” following Williamson (1990). Williamson identifies health and education as “proper objects of government expenditure” but says nothing about SP. The more explicitly poverty-focused policy discussions of the 1990s gave more attention to SP, although it was seen as a short-term palliative to help mostly middle-income countries address downside risks and Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 help specific disadvantaged groups (such as the disabled) who could not be expected to benefit directly from economic growth.6 This policy prioritization regarding SP often came alongside a view (not always explicit) that governmental efforts at reducing inequality (including SP but not confined to SP) in poor countries should wait until the countries are not so poor. Poverty reduction was perceived as the near-term goal, while economic growth in a market economy (supported by sound public investments in health and education) was seen as the primary instrument. There are signs that this view of SP has been changing. Over the last 20 years or so, there has been a greater emphasis on the development role of SP in low- and middle-income countries, as an “equal partner” with more traditional development instruments. The volume of development assistance for SP by the World Bank and most other international development agencies (both multilateral and unilateral) rose appreciably alongside the change in thinking. As a prior director in charge of the World Bank’s social protection efforts, Yu (2016), put it, “Social protection is no longer considered to be a luxury.” Nonetheless, as fig. 1 illustrates, the pattern of a rising share of GDP devoted to SP as GDP rises is still evident in recent data. (Possibly the SPEC would be even steeper if data were examined for, say, 25 years ago; the article will return to this point.) This paper asks why higher SP shares are seen among countries with higher GDP – a rising cross- country SPEC, as in fig. 1. Is it coming from how higher national income influences demand for SP, or does it stem from other (omitted) characteristics that are jointly correlated with national income and the SP share? The answer matters to how one interprets the SPEC. One interpretation could be that the rising SPEC stems from preferences – people in poor countries care more about other things than SP. The alternative view is that they care just as much (or possibly more) about SP but weak institutions, deficient infrastructure, and other conditions typical of poor countries make it harder to convert their notional SP demands into public spending outcomes – in short, that poorer countries face higher costs in effectively implementing this type of public spending. Prices of competing human development services (health and education) may also play a role, to the extent that these services are relatively expensive in richer countries, encouraging substitution toward SP transfers within public budgets. To perform the present study, a new dataset was assembled on social protection expenditure covering 142 countries in the period 1995–2020. This is almost certainly the largest dataset (in terms of country and time coverage) on social protection expenditure currently available. This dataset makes it possible to analyze the patterns of social protection expenditure in a considerably larger country sample than the samples used in similar studies in the field7 and thus represents a contribution to the literature in itself. The dataset is the first to include panel observations on SP expenditures in low- and middle-income countries. This dataset is complemented with a second new dataset that focuses on the pandemic period 2020–2021 and covers SP programs under the designated pandemic stimulus budget of each of 154 countries. 6 For example, the World Bank’s (1990) influential World Development Report, Poverty, identified a role for SP, but this was (explicitly) secondary to the report’s two-part strategy for sustained poverty reduction by combining policies to encourage labor-intensive economic growth in a market economy with investments in health and education. 7 Auteri and Costantini (2004) use a sample of 16 OECD countries in the period 1981–1998. Clemente, Marcuello and Montañes (2012) perform a similar analysis using a sample of 18 OECD countries in the period 1980–2001. Brückner, Chong, and Gradstein (2012) use a sample of 142 countries in the period 1960–2007 but focus only on total government expenditure and do not provide results on social protection expenditure. 522 Lokshin, Ravallion, and Torre The present article builds on previous work studying the relationship between country income and social expenditure. Auteri and Costantini (2004) study the elasticity between income and social expendi- ture (including health) in a sample of 16 OECD countries over the period 1981–1998 using a time-series framework. They establish the existence of a co-integrated, not spurious correlation between income and social expenditure, with the elasticity being above or below unity depending on the estimation method. This finding, which matches the pattern observed in fig. 1, is the starting point of the present study, as its Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 focus is on understanding the factors that account for the relationship between income and SP expendi- ture. The present article formulates a series of hypotheses to explain the SPEC picture. Testing these hy- potheses can be readily recognized as a classic issue in causal inference. The idea of omitted variable bias (OVB) provides a natural framework for structuring this effort. The hypotheses suggest specific country characteristics that might be expected to influence the SPEC while also being correlated with national income. (The study strives to avoid a “kitchen sink” approach to controls, but instead to carefully for- mulate hypotheses based on the literature or the study’s reasoning.) These characteristics are interpreted as reflecting differences in the implicit prices faced in providing SP; for example, the implicit price of SP is higher when government effectiveness or technology access, in general, is lower. The article does not claim that its list of potential confounders is exhaustive, but (as will be seen) it is enough to be confident about the answer to the main question posed by this paper. The OVB framework makes it possible to test these hypotheses in a straightforward way – by estimating whether the correlation between country income and SP expenditure is affected after the inclusion of additional regressors in a linear model. The specific hypotheses that are considered relate to relative prices for human development services, political economy (governmental accountability, the capacity for public service provision more broadly, and redistributive policy making), the distributional impacts of economic growth, access to appropriate technologies, population ageing, and the selection processes for observed public spending data. The study argues that some of the hypotheses under consideration can already be dismissed from what is already known, based on the literature. Among the rest, each hypothesis points to one or more potential con- founders in estimating the SPEC, which the study uses in testing for OVB. The article also points to some likely theoretical ambiguities in how the confounders can influence SP spending; for example, having a more capable government can be expected to reduce the unit cost of attaining a given level of social pro- tection (thus reducing SP spending), but also increase the desired level of protection (increasing spending). A separate analysis of SP expenditure during the pandemic is also performed using the study’s second dataset on the expenditure on SP programs within countries’ pandemic stimulus budgets. The motivation in doing this analysis is that, prior to the pandemic, the leadership of low- and middle-income countries may have wanted to see more resources devoted to SP, but faced “stickiness” in budget allocations and difficulties in domestic resource mobilization (see, for example, Alesina and Passarelli 2019). There may well have been a genuine change in thinking about SP among policy makers, but it just takes a long time for this to be evident in the SPECs. A major shock can potentially break the hysteresis and so reveal the new priorities. Such a shock can also be expected to systematically influence the composition of SP spending; for example, the pandemic encouraged a greater emphasis on job protection/retention schemes (Gentilini et al. 2021; Demirgüç-Kunt, Lokshin, and Torre 2022). The present article asks if this opportunity was taken up by poor countries, such that SP ceased to be a luxury good during the pandemic. What was found? By using the new cross-county panel dataset on SP expenditure, the study first shows that the positive income effect on SP shares evident in fig. 1 remains; country fixed effects are included, although the slope of the SPEC is attenuated considerably – in line with previous findings in the literature, such as those of Auteri and Costantini (2004), which indicated an elasticity of income to social expenditure potentially below unity. Next, the study argues that the tendency for the SP share to rise with national income is mainly attributable to OVB stemming from multiple, time-varying, confounders. When country The World Bank Economic Review 523 fixed effects are also included, the SPEC that is obtained after controlling for the confounders turns out to be negatively sloped with respect to GDP per capita. In studying the policy responses to the pandemic, the article does not find that the opportunities created for implementing a radically new policy regime (with greater emphasis on SP in poor countries) were taken up in general. Relatively low public spending on SP among poorer countries during the pandemic appears to stem mainly from weak government effectiveness in public service delivery and from younger Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 populations rather than low income per se, again rejecting the luxury good hypothesis. Section 2 outlines the series of alternative hypotheses. Section 3 describes the data assembled for this paper. Section 4 presents the results on testing whether one or more of the alternative hypotheses can explain why higher SP spending is seen as a share of GDP in richer countries. Section 5 turns to the SP responses to the pandemic. Section 6 concludes. 2. Alternative Hypotheses There is more than one way to interpret the finding that poorer countries tend to devote lower shares of their national income to SP. The relation in fig. 1 could stem from a preferred policy hierarchy that simply reflects the preferences of citizens, most of whom (it is inferred) consider SP to be a luxury good. The take-away for policy is that the relatively low level of SP spending in poor countries should not be of much concern. If citizens of poor countries would rather see public spending going elsewhere, why should anyone care? However, there are a number of alternative explanations for the rising SPEC, with different implica- tions for development policy. These can be thought of as differences in the costs incurred by governments in supplying SP, as reflected in omitted country characteristics in a static cross-country comparison, such as fig. 1. This can be addressed in part by introducing country fixed effects into a SPEC estimated on a longitudinal data set. However, there are also concerns about omitted time-varying characteristics corre- lated with GDP. The study motivates thinking about those omitted characteristics in the form of a series of hypotheses. It is noted whether the existing literature already includes what appears to be an adequate basis for rejecting the hypothesis and proceeding no further. For the rest, the study proceeds to further empirical testing. The hypotheses relate to factors that could explain the positive correlation between country income and SP expenditure. The study’s hypotheses can be grouped under four headings: (1) relative prices, (2) governance, (3) distribution, and (4) technology and demographics. The first hypothesis starts from the long-standing observation that Engel curves can shift with relative prices. While SP does not have its own “relative price” in any obvious sense, there are explicit prices for substitutes for SP within the gamut of public spending on human development. Given that health and educational services are more costly in richer countries, it would be expected to see substitution in favor of SP within public budgets8 . This leads to: Prices: The relative prices hypothesis. Higher relative prices of competing human development services in richer countries lead their governments to switch to SP transfers. The next set of hypotheses relates to governance and political economy, starting with: Governance 1: The accountability hypothesis. SP is equally important for people from poor and rich countries, but the governments of poorer countries tend to be less responsive to their citizens’ demands. An autocratic dictator likely faces fewer incentives to respond to citizens’ demands (at least once the local elite is protected) than a government in a robust democracy. Even among democracies, some have better institutions (such as through greater freedom for the media) to help ensure accountability. 8 For 2017, the price data (described later) indicate that the price of health services relative to food and non-alcoholic beverages has a correlation coefficient of 0.61 with (log) GDP per capita; for education services the correlation coefficient is 0.62 (The supplementary online appendix provides the scatter plots, fig. S1.1). 524 Lokshin, Ravallion, and Torre Figure 2. Government effectiveness and GDP per capita. 2.5 SGP FIN CHE NOR CAN NZL LUX JPN AUS Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 1.5 GBRAUT ISL USA ISR Govenment effectivness ranking IRL LVA LTU CZE GEO URY QAT 0.5 JAM CHN SYC HRV ITA THA SAU CPV RWA IND PHL SRB TTO 0.0 ARM ARG MAR MEX MNG UKR DOM −0.5 KHM ETH STPPAK BFA LAO TLS MLI MMR LBR GNQ −1.5 COD SDN Correlation = .86 SOM −2.5 6.5 7 8 9 10 11 12 Log of GDP per capita (2019 PPP) Source: Government effectiveness data from World Bank (2022a). GDP data from WDI. Note: The graph plots, for the 154 countries with the required data, the 2019 WGI government effectiveness index (vertical axis) against the log 2019 GDP per capita, in USD PPP prices (horizontal axis). Country codes are the U.N.’s Alpha 3 list.9 The literature has provided mixed evidence related to the accountability hypothesis. Lindert (1994) argued that the rise of democracy, alongside economic development, was an important factor in the in- crease in SP spending in today’s rich world – an increase that took place at a time when it was not nearly so rich. However, Mulligan, Gil, and Sala-i-Martin (2004) find little sign that democracies tend to pursue different social policies, once one controls for other economic and demographic differences. Dincecco (2009) argues that more absolutist (and presumably less accountable) political regimes in pre–World War 1 Europe tended to raise less revenue generally, which would have constrained SP spending. Governance 2: The government effectiveness hypothesis. Poorer countries have less-capable govern- ments for delivering public spending, whether on SP or something else. The literature on institutions and development has often pointed to ways in which being a richer country can promote better institutions and polices.10 The cost of delivering a unit of social protection is presumably lower in countries with better institutions. That can explain Ravallion’s (2017) finding that the coverage of social protection is positively correlated with national income. Government effectiveness is likely to be correlated with GDP, but it also has its own independent variation, in that there are poor countries with relatively effective gov- 9 “Country Codes List,” https://www.nationsonline.org/oneworld/country_code_list.htm#R. 10 Thus, this strand of the literature gives much attention to the likely endogeneity of institutions in explaining differences in rates of economic growth; see, for example, Acemoglu, Johnson, and Robinson (2005). The World Bank Economic Review 525 ernments, and some rich countries for which one could not reasonably say that. This is illustrated by fig. 2, which plots the “government effectiveness (GE) index” from the World Bank’s Worldwide Governance Indicators (2023) (WGI) against GDP per capita. The GE index is highly correlated with GDP per capita. However, there are low-income and middle-income countries with relatively high GE values and high- income countries with low values. (For example, Saudi Arabia’s GE index is about the same as Rwanda’s.) The last of the three hypotheses under the governance heading concerns local resource mobilization: Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Governance 3: The domestic resource mobilization hypothesis. Poorer countries are more constrained in raising revenues needed for financing SP spending.11 The key assumption here it that the easier it is for the government to mobilize resources domestically through taxation, the easier it will be to finance transfers as SP. By contrast, most external funding sources (such as commercial and development banks) are more demanding that the monies go toward things that are traditionally seen as investments, with a pecuniary return that would help assure loan repayment. (When available, development grants appear more amenable to spending on SP.) Governance 2 and 3 hypotheses clearly overlap. The public services that a government is able to provide depend on its ability to finance them through domestic taxation, though also through borrowing or grants, including development assistance. As noted, the accountability of governments might also be expected to influence resource mobilization. One might subsume the hypotheses Governance 3 within Governance 2, but it is likely to be of sufficient importance to warrant its own reckoning as a covariate of SP spending. Next, three hypotheses related to the distribution of income are considered. Distribution 1: The median-voter hypothesis. More skewed income distributions in rich countries en- courage spending on SP as a form of redistribution. The premise here is that more (positively) skewed income distributions encourage more redistributive public spending as formalized in the famous median- voter theorem (Meltzer and Richard 1981). Social protection is, to some degree, a redistributive inter- vention. Skewness in the income or wealth distribution can then be expected to matter to the political economy of SP spending. Additionally, if poorer countries tend to have less-skewed income distributions, then one might expect to find that SP spending is a lower priority in those countries. There are reasons to question this hypothesis. One reason is that some (mostly developing) countries are not democracies, so the voting mechanism postulated by the median-voter theorem is absent. However, it is not clear that this would fully neutralize the relevance of income distribution; even dictators cannot entirely ignore their citizens’ wishes.12 A more compelling reason to question this hypothesis is that the evidence suggests that poorer countries tend to have more unequal distributions.13 The (small) literature on the role of distribution in influencing public spending has almost solely used the Gini index (as in, for example, Shelton 2007), rather than skewness, but the two are likely to be highly correlated. Empirically, skewness is a more common fea- ture of income distributions in low-income countries than in high-income ones; in fact, the correlation coefficient between (log) GDP per capita and the ratio of the mean to the median incomes is −0.405.14 Thus, if anything, the political economy of redistribution (or at least the median-voter version) implies more redistributive effort in poorer countries. So, it appears that enough is known already to reject this hypothesis as an explanation of the rising SPEC. It will not be tested further. 11 This can be seen as an instance of the more general argument that the capacity for domestic taxation is important for development outcomes, as in Besley and Persson (2011). 12 Han (2021) finds that in authoritarian regimes that hold elections, redistributive policies are implemented in the run-up to elections even when election results are predetermined. 13 See, for example, World Bank (2006) and Milanovic (2016). 14 The analysis measures the ratio of the mean to the median, with both measured from the household surveys, as processed in the World Bank’s PovcalNet site. While the ratio of the mean to the median is not strictly skewness (the third moment of the distribution), it is an intuitively appealing indicator. The ratio exceeds unity in all countries, as one would expect. 526 Lokshin, Ravallion, and Torre Distribution 2: The unequal growth hypothesis. Economic growth in market economies tends to in- crease inequality. This leads to political demands for redistributive SP policies. The idea that higher in- equality creates demands for redistribution is familiar from the political economy literature (including, for example, Alesina and Rodrik 1994). In the present context, the concern is that the positive income effect on the SPEC is actually picking up the effect of economic growth on inequality, leading to higher SP spending. Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 However, in light of accumulated evidence from household survey data, the basic premise of this hy- pothesis can be questioned. Empirically, a far more mixed picture is seen, with inequality decreasing roughly as often as it increases in growing economies.15 These findings have been mainly from research on developing countries. However, a similar result is found when one includes high-income countries; in the global data set that is used below, the correlation coefficient between the first difference of the log Gini index and the first difference of log GDP per capita is −0.09.16 So, from what is already known, this hypothesis can also be dismissed. One goal of SP is to ensure some minimum standard of living in a society, as typically indicated by a poverty line. Naturally, national SP efforts depend on prevailing national poverty lines rather than on international absolute lines, which aim to have common real value across countries (such as the World Bank’s $1.90-a-day line). This motivates the next hypothesis: Distribution 3: The relative poverty hypothesis. Higher national poverty lines in rich countries imply that more public SP spending is needed to reduce poverty. In assessing the a priori case for this hypoth- esis, it must first be recognized that there are two, potentially offsetting, ways in which a higher average income will alter the cost of reducing poverty using SP spending. Poverty lines are typically relative in rich countries, meaning that they rise in proportion to the mean or median income; low- and middle-income countries, by contrast, tend to have absolute or only weakly relative poverty lines, meaning that they have an elasticity to the mean that is less than unity (Ravallion, Chen, and Sangraula 2009). In contrast, the national poverty gap index (PG) using weakly relative poverty measures is likely to be a decreasing func- tion of mean income.17 It is, then, an empirical question whether the aggregate poverty gap per capita (PG times the national poverty line) is increasing or decreasing with mean income. If the (positive) elasticity of the national poverty line to mean income exceeds (in absolute value) the negative elasticity of the national poverty gap index with respect to mean income, then the cost of targeted efforts in using SP to reduce poverty will tend to be higher in rich countries. To test this, the analysis draws on estimates from the literature of these two elasticities. Ravallion (2016b, Appendix) estimates the cross-country elasticity of the national poverty line to the mean to be 0.52 (robust s.e. = 0.04; n = 598).18 Using the estimates of the relative poverty gap index at the country level (calibrated to predicted national poverty lines) obtained by Ravallion and Chen (2019), and also allowing for country fixed effects, the analysis obtains an elasticity with respect to the mean income or consumption of −0.56 (s.e. = 0.08; n = 144). Thus, the two effects of higher average income on the national poverty gap per capita (PG times the poverty line) – one on the national poverty line and one on the PG itself – essentially cancel out, suggesting that the income elasticity of the cost of eliminating poverty is likely to be close to zero. 15 Ravallion (2016a, chapter 8) reviews the evidence on this point. Note that the reference here is to relative inequality; absolute inequality tends to rise with growth; on this distinction and the evidence, see Ravallion (2021). 16 For the levels regression with country fixed effects the partial correlation coefficient is −0.03. 17 PG is the aggregate proportionate distance below the poverty line (expressed as a proportion of the line and counted as zero for those above the line) per capita of the total population. 18 This is based on the compilation of national poverty lines across countries and over time from Jolliffe and Prydz (2016). Ravallion (2016b, Appendix) regressed the log of the national poverty line on the log of mean income including country fixed effects. The World Bank Economic Review 527 In short, based on the existing evidence, the relative poverty hypothesis is not considered further, since the cost of eliminating poverty is not expected to be correlated with mean income. A further hypothesis relates to access to technology for supporting SP programs: Technology: The ICT hypothesis. The governments of poorer countries have less access to informa- tion and communications technology (ICT), which makes SP programs more costly options for public spending. The literature on SP has not often discussed this issue, but it is a seemingly plausible assump- Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 tion that SP programs tend to be ICT-intensive, such as in creating and maintaining social registries of participants and payment and monitoring systems. There is evidence that such ICT capabilities enhance the quality and coverage of SP programs in poor countries.19 Better communication infrastructure in a country can facilitate new options for using mobile money technologies that have proven to be helpful for cash transfer delivery in some low-income settings.20 However, with limited access to ICT at the national level due to poor infrastructure, the cost of implementing SP programs can be expected to be higher in poorer countries, leading policy makers to substitute SP in favor of other types of public spending. This hypothesis will be tested using ICT usage indicators at the national level over time. The next hypothesis concerns population aging: Demographics: The aging hypothesis. Richer countries tend to have older populations, which increases the demand for SP. The fact that richer countries tend to have more elderly populations follows almost immediately from the well-known Preston (1975) curve, whereby life expectancy at birth tends to be an increasing (concave) function of average income across countries.21 With an older population in richer countries, demands can be expected for extra public spending to help support the elderly, who have diminished capabilities for supporting themselves; Shelton (2007) finds cross-country evidence to support this claim. Since pensions are one component of SP spending, countries with an older population will tend to have higher SP spending, though non-pension components may also be higher. A preliminary look at the evolution of SPECs excluding pension spending (supplementary online appendix figs. S1.3 and S1.4) does not suggest that pension spending explains the pattern in fig. 1. The aging hypothesis will be tested. Lastly, in the context of the pandemic, a further hypothesis is considered: The COVID hypothesis: Richer countries were impacted more by the pandemic and responded accord- ingly with their own SP spending. In this view, rich countries had more of a shock to protect themselves from, reflecting higher levels of social and economic interaction that helped spread the infection (prior to vaccine availability). COVID mortality data can be used to measure the severity of the pandemic shock.22 These data do suggest a greater COVID impact in countries with higher GDP per capita; in the present data set, the correlation coefficient between COVID deaths per capita and GDP per capita is 0.439. The analysis tests this as an extra covariate for SP spending during the pandemic. 3. Data and Descriptive Statistics The primary dependent variable in the analysis is the share of national income devoted to public SP expenditure. One of the main contributions of this study is assembling the largest dataset on public SP expenditure currently available. In this dataset, the coverage is across countries and over time up to (but not including) the pandemic. This is complemented with a second dataset, which includes the designated SP responses to the pandemic in the period 2020–2021. 19 For examples of the use of ICT in registration, payments and other functions, see World Bank (2022b). 20 See, for example, Jack, Ray, and Suri (2013). 21 For more recent descriptions and analysis, see (inter alia) Gómez and Hernández de Cos (2008) and Ritchie and Roser (2019). 22 It is acknowledged that these data may be subject to measurement error, especially in many low-income countries. Another indicator of the severity of the pandemic is the change in GDP during the pandemic, but this is endogenous to the SP efforts in response. 528 Lokshin, Ravallion, and Torre Table 1. Share of SP Spending as Percent of GDP by Year Stratified by 1995 GDP per Capita. GDP per capita by 1995 quintiles 1995 2000 2005 2010 2015 2019 1 (lowest) 1.37 1.68 2.47 3.26 3.52 4.20 2 3.49 6.64 5.19 6.36 5.74 6.01 3 4.57 4.90 4.84 6.75 6.18 6.83 4 9.59 10.00 9.64 10.97 10.24 10.37 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 5 14.16 12.79 13.17 14.92 15.02 14.72 Total 8.15 8.35 8.07 8.84 8.48 8.91 Source: Calculations based on social protection expenditure data from the authors’ data set. Pre-Pandemic Data on Social Protection Data on public spending at the country level were assembled for the purpose of this paper from the records kept by multiple agencies, notably Eurostat, OECD, ECLAC, IMF, and the World Bank. All these sources have been drawn upon to provide as complete a data set as feasible. The appendix provides details on the specific sources for every country and year, covering 142 countries from 1995 to 2020 (table A1.1). This seems to be this is the largest dataset (in terms of country and time coverage) on social protection expenditure currently available. There are limits to how far cross-country comparisons, with limited degrees of freedom, can be con- sidered conclusive about the determinants of the SPEC. However, by pooling cross-country observations with time series data, there are 2,481 observations for assessing the relative importance of these various factors. On average, there are 17.5 time-series observations per country with non-missing shares of SP expenditure. The analysis focuses mainly on total SP spending, though the study notes any important differences between pensions and other SP spending, since public pension schemes often appear to be serving a somewhat different role from other types of SP spending (notably the more poverty-focused category of “social assistance”). The sources make it possible to distinguish pension from non-pension SP spending for more than 1,500 observations across 104 countries, although coverage is less complete for low-income countries. The panel data are unbalanced. A saturated panel would have 3692 ( = 142 × 26) observations. So, there is about two-thirds (0.67 = 2481/3692) coverage for the basic SPEC (though this falls further when covariates are introduced). The number of country observations per year is increasing, from 75 in 1995 to a maximum of 112 in 2017. In no year is there SP spending data for all 142 countries; 45 countries have 26 years of data, and 73 have 20 or more years. The supplementary online appendix provides a series of graphs (fig. S1.2) that show the missing observations are more likely to be from lower-income countries. (A similar pattern is found for most other covariates discussed below.) Descriptive Statistics on SP Spending Some simple descriptive statistics based on the (pre-pandemic) panel data set are instructive. On average, the study finds that SP spending has risen over time as a share of GDP, and that this has been more pronounced among initially low- and middle-income countries. This can be seen in table 1, which tabulates the SP spending share by countries classified according to their initial (1995) GDP per capita. For the lowest quintile of countries, the share of GDP going to SP increased by a factor of three between 1995 and 2015, from 1.4 percent to 4.2 percent.23 By contrast, the share has been quite stable over time for the upper two quintiles. 23 Much of the rise in SP spending by initially poorer countries was coming from public outlays on pensions; details are provided in table S1.1 in the supplementary online appendix. The World Bank Economic Review 529 Figure 3. Nonparametric SPECs for various years. 20 Spending on social protection as percent of GDP 1995 2010 2019 2000 2005 2015 15 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 10 2019 2015 5 2010 2005 2000 1995 0 6 7 8 9 10 11 Log GDP per capita Source: Analysis based on social protection expenditure data is from the authors’ data set. GDP data from WDI. Note: The nonparametric regression lines are smoothed scatter plots (using the lowess command in Stata). Each line represents the data for the specific year indicated. The five richest countries in the income distribution of every year are excluded from each Lowess estimation. The rise in the SP share among initially poorer countries that is seen in table 1 could reflect either economic growth in those countries or a change in policy priorities, whereby the share rose at a given level of GDP per capita, implying an upward shift in the SPEC at the lower end. That can be assessed using fig. 3, which augments fig. 1 to provide the nonparametric SPECs for various years. If anything, the SPECs have tended to shift downwards over time. Among low-income countries, little or no change is seen in the SP share at a given level of GDP per capita except for a drop in average levels between 1995 and 2000. In the middle–upper income range, there is actually a decline after 1995. Across all levels of GDP, the 1995 SPEC is unambiguously higher than those for 2005, 2010, 2015, and 2019, while the 2019 SPEC is unambiguously lower than those for all years except 2000 and 2005. The greater emphasis on SP in development policy discussions in the new millennium does not appear to have shifted the SPEC upwards. This finding is not inconsistent with the recent expansion in SP programs in developing countries; it is just that these programs have typically had low coverage (on average) in low-income countries, as documented in Ravallion (2017). The rise in SP spending as a share of GDP in developing countries is coming from higher GDP rather than from a change in development priorities in favor of a higher share of national income going to SP at given GDP. Data on Social Protection Spending during the Pandemic The panel data set on SP spending prior to the pandemic is augmented by a database compiled by a team including two of the present authors and documented in Demirgüç-Kunt, Lokshin, and Torre (2022). This provides estimates of the expenditures incurred in social protection measures that the government imple- mented in response to the pandemic over 2020 and 2021 in 154 countries. The data draw on information from the Global Database on Social Protection Responses to COVID-19 (Gentilini et al. 2021), budget 530 Lokshin, Ravallion, and Torre data from official documents (including IMF Article IV revisions and other international organizations’ related documents), government websites, and news sources. The key criterion for the inclusion is that the component of SP spending must have been designated as a response to the pandemic. In some cases these were new programs while others were expansions to existing programs. To the extent feasible, the study estimated the budget corresponding to the expansion of an existing program (for example, the cost of extending the eligibility period of unemployment benefits) and not the whole pre-existing program, Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 though it is not always possible to be certain of this since sometimes the source was ambiguous about this difference.24 Data on the Covariates The available data limit the study’s ability to test the Relative Prices Hypothesis. The International Comparison Program provides estimates from each ICP round of the prices of different consumption cate- gories, including health services, education services, and the aggregate category of food and non-alcoholic beverages. With this information, the relative price of health and education to food and non-alcoholic bev- erages can be easily calculated.25 However, the ICP data are only available for the years 2005, 2011, and 2017.26 The health and education price relatives are highly correlated (r = 0.92 for 2005, r = 0.90 for 2011, and r = 0.82 for 2017); results are shown with them entering separately as well as jointly. In testing the Governance 1 hypothesis, the analysis uses measures of the responsiveness of governments to their citizens: namely, their “voice and accountability” (VA) index from Word Bank’s WGI. The index is a composite of indicators on democracy, electoral processes, accountability of public officials, rights, reliability of government budget documents, transparency in policy making, the freedoms allowed for citizens, and citizens’ trust in their government. The index comes normalized to have mean zero and a standard deviation of unity in every year. Two other measures of accountability are considered. The first is the electoral democracy index (also known as the polyarchy index) produced by the Varieties of Democracy (V-DEM) project (Coppedge et al. 2023). This measures the extent to which a country has free and fair electoral competition, suffrage is extensive, political and civil society organizations operate freely, there is freedom of expression, and the media are independent. The index ranges from 0 to 1, where 0 indicates a regime furthest away from electoral democracy and 1 is a full electoral democracy. The second measure is the political competition index produced by the Polity5 project (Center for Systemic Peace 2020). This categorical index measures the degree of political competition in a given country, independently of the political regime, and ranges from a value of 1 (meaning that political competition is suppressed) to a value of 10 (meaning that political competition is institutionalized and electoral). This measure is used as an additional test of the Accountability Hypothesis under the assumption that regimes where political competition is higher tend to be more accountable and responsive to citizens’ demands. In testing the Governance 2 hypothesis, the study again turns to the WGI, which also provides a mea- sure of government effectiveness (GE) across countries (World Bank 2022a). GE is a composite index of indicators of the quality, coverage, and citizen satisfaction with public goods and services (roads, public transport, electricity, health and education services, water and sanitation, bureaucratic quality). Similar to VA, the GE index is normalized to have mean zero and a standard deviation of unity. 24 The study checked against the IMF estimates on the total additional public spending (whether on SP or something else) implemented during the pandemic (IMF 2021); in no case did this study’s estimate exceed the corresponding IMF number. 25 The ICP provides information of the price level of each category indexed to the world average, which takes a value of 100. The ratio of these price levels provides an indicator of relative prices. 26 The number of countries for which social protection and ICP data are available is 81 in 2005, 99 in 2011, and 102 in 2017. The World Bank Economic Review 531 For the Technology Hypothesis, the study draws on data from the International Telecommunications Union (ITU) that focus on the use of ICT.27 The analysis focuses on two technologies, namely mobile phones and the internet. For the former, the number of mobile phone subscriptions per capita is used. For the latter, the percentage of the population using the internet is used. Both these measures are only available from the year 2000 and onwards. While there are about 2,500 observations of the expenditure on SP by country and year, the covariates Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 used to test the hypotheses laid out in section 2 sometimes have a shorter time period and more limited geographical coverage. The subsample for which all variables except relative prices are available includes about 1,500 observations covering 107 countries. Relative prices are only available for three years (2005, 2011, and 2017), and therefore, the subsample including them (in addition to the remaining covariates) is limited to 220 observations from 95 countries. 4. Tests of the Rival Hypotheses for Why a Rising SPEC Has Been Observed The observations in “Descriptive Statistics on SP Spending” in section 3 point to the role of economic growth in fostering higher SP spending in developing countries. This requires a rising SPEC, whereby shares of GDP devoted to SP tend to be lower in countries with lower GDP, as in fig. 1. But the main question still remains: Why is the rising SPEC being observed? It is well-recognized that cross-country comparisons, such as in fig. 1, can be deceptive, given that country characteristics could be correlated with national income and also influence SP spending. If one has not controlled for those characteristics, then GDP is endogenous in the estimation of the SPEC. The study addresses this concern by adding controls for (time-varying) observables and using the panel struc- ture of the data to deal with (time-invariant) unobservables, assuming that the correlation between the error term of the regression model for the SP share and the regressors is fully captured by an additive country fixed effect.28 Under this assumption, the analysis will estimate the SPEC by pooling years and countries, and including country fixed effects, thus relying solely on the inter-temporal variances. The study acknowledges that biases may remain due to time-varying unobservables (or measurement errors). The textbook solution is to use one or more instrumental variables (IVs), but the study is skeptical of the possibility of finding valid IVs in this context.29 Careful selection of the control variables appears to be the best option. In doing so, it is necessary to be cognizant of the “bad control” problem (e.g., Angrist and Pischke 2009), a situation when potential controls are blocking the path of the true effect of GDP on SP expenditures. If bad controls are correlated with the regressor of interest, including these in the regression might result in smaller magnitude coeffi- cients. It is difficult to assess whether the controls being used are good or bad, as the causality could go in both directions. However, the study argues that most of the indicators it uses to test out hypotheses, for example, that government effectiveness, voice and accountability, and the levels of technology penetration, are determined by longer-term processes and react to external shocks in a much slower way compared to 27 One might prefer to test this hypothesis with information on the use of ICT services by each country’s government, but there is no data source that provides enough time and geographical coverage on that dimension. The study also acknowledges the potential limitation of this indicator as different state arrangements may distribute and use ICT technology in SP administration differently, even if the population have the same levels of usage. 28 It is expected that any reverse causality, whereby GDP responds (positively or negatively) to SP spending, is a longer-term effect, such that the short-term changes in GDP (after allowing for time-varying controls and country fixed effect in the panel data set) can be treated as conditionally exogenous. 29 Brückner, Chong, and Gradstein (2012) use oil price shocks as the IV for GDP in testing Wagner’s law using Engel curves for public spending. However, in the present context, oil price shocks could induce higher SP spending by altering the distribution of income (invalidating the IV). Instead, this study considers ways in which income distribution can be a potential confounding variable in its own right. 532 Lokshin, Ravallion, and Torre Table 2. Summary of Hypotheses for Explaining the Rising SPEC. Further testing? Control variable Prices Yes Relative prices for health and education services Governance 1: Accountability Yes Polyarchy index Accountability Yes Political competition index Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Accountability Yes Voice and accountability index 2: Government effectiveness Yes Government effectiveness index 3: Resource mobilization Yes Tax revenue as a share of GDP Distribution 1: Median voter No 2: Unequal growth No 3: Relative poverty No Technology Yes Mobile phone suscriptions Yes Internet usage Demographics Yes Share of the population age 65+ GDP per capita. It could be argued that these controls were determined prior to the variable of interest and, as such, cannot themselves be the outcomes of the causal nexus. Section 2 pointed to a number of potentially confounding variables correlated with levels of GDP (or for which changes over time are correlated with growth rates). Table 2 summarizes the various hypotheses, including the control variables (when further testing appears to be warranted based on what is known already from the literature). Using the OVB analysis framework, the analysis builds up its tests by adding control variables to the basic SPEC in which the SP spending share depends solely on GDP per capita. The analysis initially adds each candidate variable one at a time, to see how far each one can explain, on its own, the income effect on the SP share. The expectation, stemming from the OVB framework, is that the potential confounder will reduce the coefficient on (log) GDP per capita and possibly eliminate its effect. The analysis then tests an encompassing model, recognizing that these variables are correlated with each other. Regression Specifications The following linear regression model of the SPEC is used to test each hypothesis separately: SPit = α + β lnGDPit + γ Xit + εit (i = 1, . . . , N; t = 1, . . . , T ) (1) GDPit Here SPit is social protection spending in country i at date t; GDPit is GDP per capita; and Xit is the control SPit variable implied by each hypothesis. The income elasticity of demand for SP is then given by 1 + β/( GDP it ); an upward sloping SPEC (β > 0) implies an income elasticity exceeding unity. The analysis considers two assumptions about the error term, εit , in equation (1). For the first, it is taken to be a standard innovation error term, orthogonal to the regressors, and to selection into the panel data set, while in the second, the analysis includes country fixed effects, potentially correlated with the regressors and with selection. The analysis also estimates a more general (encompassing) specification that generalizes (1) in two respects. First, given that the controls are also correlated with each other, the encompassing model includes all the controls so they can fight it out as to which is more important. (When there is more than one for a given hypothesis, the analysis selects what appears to be the best representative.) Second, the study allows for any (continuous) nonlinearity in the SPEC. Such nonlinearity in the national income effect on SP spending shares (as evident in fig. 1) may well be confounding in this context. (For example, if the true relationship is quadratic, but the squared value is excluded, then a control variable may simply pick The World Bank Economic Review 533 up this omitted variable.) Combining these assumptions, the encompassing regression has the following form: SPit = f lnGDPit + X it γ + ηi + νit (i = 1, . . . , N; t = 1, . . . , T ) (2) GDPit Here f (. ) is some (data-determined) smooth nonparametric function and X it is now a vector of control variables. (The income elasticity of demand for SP is then 1 + ∂ ∂ ln f (. ) .) The study estimates (2) as a partial Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 lnGDPit linear regression (Yatchew 1998) using Stata’s PLREG routine (Lokshin 2006). Results As can be seen from the “baseline” panel in table 3, the slope coefficient in the simplest linear regression specification for the SPEC gives a semi-elasticity of 3.78 (implying a full elasticity of SP spending to GDP of 1.42, when evaluated at the mean SP share for 2019). This falls to a semi-elasticity of 1.37 (full elasticity in 2019 of 1.15) when the country effects are included.30 Qualitatively, this is what one would expect if the omitted country characteristics in the SPEC based on cross-country comparisons (such as fig. 1) tend to be jointly positively correlated with the SP share and GDP. Nonetheless, a positive and significant income effect on the SP share of national income remains when the analysis allows for the latent country effects.31 The study still finds a statistically significant nonlinear income effect on the SP shares when the anal- ysis allows for country fixed effects in the PLREG estimator (the “baseline” panel in table 3). Figure 4 compares the estimated nonparametric subfunctions, fˆ (lnGDPit ), with and without the country effects, where the fˆ (lnGDPit ) values controlling for the country effects are centered on the overall mean of the SPit country-specific means of GDP it . The positive slope is evident either way, but is attenuated considerably by introducing the country effects. Comparing the top quintile of countries ranked by GDP per capita with the bottom quintile, the mean SP shares of national income are 9 percent and 5 percent with country fixed effects, as compared to 13 percent and 2 percent without them. Next, the study tests the effects of adding time-varying country characteristics, corresponding to the various hypotheses laid out in sections 2 and 3. The results of testing the hypotheses one at a time (equation 1) are also found in table 3. This uses total SP spending. The supplementary online appendix provides corresponding results obtained by dropping public spending on pensions (table S1.3). The main results are robust to this change, though with some (unsurprising) differences in the coefficients, notably that the coefficient on the population share over 65 is much reduced when one excludes spending on pensions. In addition to standard tests, the study provides the test described in Lokshin (2006) for whether the nonparametric subfunction estimated using PLREG is significantly different from a constant (that is, no income effect). Excluding the country fixed effects, one could claim empirical support in table 3 for each of the hy- potheses outlined in section 2. Adding the country fixed effects (thus identifying the effects solely from the intertemporal variances) still leaves support for the Prices, Accountability, and Technology Hypotheses. Support for the Governance and Aging Hypotheses is diminished with fixed effects. In most cases, adding any of the postulated covariates reduces the GDP slope of the OLS SPEC, though the effect remains statistically significant. Significant income effects in the PLREG estimates are also indi- cated, in general. 30 The elasticity estimates do not differ substantially from those of Auteri and Costantini (2004) in their study on OECD countries – where they find an estimated elasticity of social expenditure (including health) to GDP of between 0.837 and 1.139. 31 The sample consists of 142 countries (N), with an average country having 17 years (T) of panel observations. As such, this panel can be considered to be “wide” (N>T), and it is not necessary to worry about the stationarity of the variables in estimating the FE model (Wooldridge 2001). Table 3. Separate Tests Using Pre-Pandemic Pooled SP Data. 534 Hypothesis Estimation method Log GDP per capita Test on f (lnGDPit ) Control variable Number of observations Coefficient Std. error Value Coefficient Std. error Baseline OLS 3.778∗∗∗ 0.499 2,451 PLREG 32.360 2,450 OLS fixed effect 1.394∗∗∗ 0.254 2,451 PLREG fixed effect 5.375 2,450 ∗∗∗ Prices: Education OLS 2.465 0.571 4.239∗∗∗ 1.399 2,282 Education/Food price ratio PLREG 5.997 4.078∗∗∗ 1.045 2,281 OLS fixed effect 1.847∗∗∗ 0.469 0.158 0.593 2,282 PLREG fixed effect 9.213 1.576∗∗ 0.740 2,281 ∗∗∗ Prices: Health OLS 2.358 0.591 7.272∗∗∗ 2.199 2,282 Health/Food price ratio PLREG 4.651 4.640∗∗ 1.859 2,281 OLS fixed effect 1.703∗∗∗ 0.476 1.991∗ 1.017 2,282 PLREG fixed effect 8.182 2.220∗ 1.231 2,281 Governance: Accountability 1 OLS 2.917∗∗∗ 0.478 9.686∗∗∗ 2.116 2,342 Polyarchy index PLREG 17.065 7.948∗∗∗ 0.616 2,341 OLS fixed effect 1.402∗∗∗ 0.259 − 0.845 1.282 2,342 PLREG fixed effect 5.451 − 1.116 0.761 2,341 ∗∗∗ Governance: Accountability 2 OLS 3.760 0.469 0.712∗∗∗ 0.208 2,189 Political competition index PLREG 29.297 0.505∗∗∗ 0.055 2,188 OLS fixed effect 1.454∗∗∗ 0.267 0.060 0.046 2,189 PLREG fixed effect 6.020 0.055 0.054 2,188 ∗∗∗ Governance: Accountability 3 OLS 2.360 0.557 2.788∗∗∗ 0.655 2,091 Voice and Accountability index PLREG 10.336 1.978∗∗∗ 0.204 2,090 OLS fixed effect 1.412∗∗∗ 0.251 − 0.641∗ 0.374 2,091 PLREG fixed effect 5.185 − 0.860∗∗∗ 0.287 2,090 ∗∗∗ Governance: Govt. effectiveness OLS 1.718 0.653 2.725∗∗∗ 0.829 2,082 Government effectiveness index PLREG 7.041 1.823∗∗∗ 0.275 2,081 OLS fixed effect 1.498∗∗∗ 0.261 − 0.569 0.431 2,082 PLREG fixed effect 5.155 − 0.512∗ 0.278 2,081 Governance: Dom. resources OLS 3.130∗∗∗ 0.582 0.288∗∗∗ 0.069 2,039 Tax revenues over GDP PLREG 19.076 0.272∗∗∗ 0.090 2,038 OLS fixed effect 1.237∗∗∗ 0.280 − 0.004 0.035 2,039 PLREG fixed effect 3.866 0.018 0.019 2,038 Lokshin, Ravallion, and Torre Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Table 3. Continued Hypothesis Estimation method Log GDP per capita Test on f (lnGDPit ) Control variable Number of observations The World Bank Economic Review Coefficient Std. error Value Coefficient Std. error Technology 1 OLS 4.269∗∗∗ 0.564 − 0.014 0.009 2,053 Mobile phone suscriptions PLREG 30.470 − 0.006 0.004 2,052 per 100 inhabitants OLS fixed effect 0.694 0.518 0.010∗∗ 0.004 2,053 PLREG fixed effect 4.325 0.013∗∗ 0.002 2,052 ∗∗∗ Technology 2 OLS 2.375 0.718 0.063∗∗∗ 0.018 2,015 Percentage of population PLREG 11.346 0.035∗∗∗ 0.008 2,014 that uses the internet OLS fixed effect − 0.480 0.653 0.038∗∗∗ 0.010 2,015 PLREG fixed effect 3.451 0.039∗∗∗ 0.005 2,014 Demographics OLS 1.220∗∗∗ 0.424 0.790∗∗∗ 0.069 2,329 Share of the population PLREG 4.816 0.721∗∗∗ 0.024 2,328 age 65+ OLS fixed effect 1.382∗∗∗ 0.262 0.027 0.037 2,329 PLREG fixed effect 5.174 0.001 0.042 2,328 Source: Analysis based on social protection expenditure data from the authors’ data set. Note: Standard errors are clustered at the country level. ∗∗∗ indicates that the coefficient is significant at 1 percent level, ∗∗ at 5 percent level, ∗ at 10 percent level. All subfunction tests are significant at p<0.001 level. The correction for selective reporting uses a cubic function of the p-score. 535 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 536 Lokshin, Ravallion, and Torre Figure 4. Nonparametric SPECs Pooling Countries and Years with and without Country Fixed Effects. 20 Lowess (full sample; no controls) Lowess (excl. five countries with highest GDP per capita in 2019) PLREG (full sample; controlling for country fixed effects) Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 SP expenditure, % of GDP (all years) 95% confidence interval 15 10 5 0 6.5 7 8 9 10 11 12 Log of GDP per capita (USD PPP) Source: Analysis based on social protection expenditure data is from the authors’ data set. GDP data from WDI. Note: The graph plots social protection spending as a percent of the GDP (vertical axis) against the log GDP per capita, in USD PPP prices (horizontal axis) for the pooled sample of 142 countries over the period from 1995 to 2020. The solid black line is a smoothed scatter plot (using the lowess command in Stata) including all the countries and the short-dash line excludes the five richest countries as measured by their log GDP per capita in 2019 (Macao, Luxembourg, Singapore, Kuwait, and Ireland). The (nonparametric) regression line in long dashes includes all countries and controls for country fixed effects. The line is centered on the country mean of the SP share. The study turns now to the encompassing test (equation 2) in table 4. Since the sample is very different (including smaller) when relative prices are included, these are left out for table 4, but the supplementary online appendix includes the regressions with relative prices (table S1.4). The results are similar. In both the linear specifications estimated by OLS (with and without fixed effects), the multiple co- ˆ on log GDP in the linear regression is much reduced variates flatten the SPEC, that is, the coefficient β and no longer significantly different from 0 implying an income elasticity of unity. (table 4 is for total SP spending; table S1.5 provides the corresponding regressions when one drops pensions; the results are similar.) However, when nonlinearity is allowed for, using PLREG, a significant income effect is still found. There is support for the Accountability Hypothesis in the specification without country fixed effects, but this switches sign when those effects are introduced; time periods in which accountability worsened saw rising SP spending, possibly to compensate. There is also support for the Domestic Resource Mobilization and the Aging Hypotheses without country fixed effects, but this is not robust to allowing for the coun- try effects. The only covariate that remains strong statistically across different specifications in table 4 is internet usage. However, it is important to warn against too much emphasis on any one factor, given the obvious sensitivity of coefficients and their standard errors to the data and econometric specification.32 32 As an alternative to the encompassing regression, the analysis uses lasso models that use regularization to penalize overfitting to get the model to prioritize the most important covariates (e.g., Belloni, Chernozhukov, and Hansen 2014). The World Bank Economic Review 537 Table 4. Encompassing Model Using the Pre-Pandemic Dataset. Linear OLS PLREG Linear OLS PLREG Country fixed effects No No Yes Yes Log GDP per capita 0.659 − 0.987 (0.918) (0.805) Voice and accountability index 0.694 0.490∗ − 0.191 − 0.553 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 (0.807) (0.250) (0.404) (0.354) Government effectiveness index − 0.655 − 1.330∗∗∗ − 0.574 − 0.221 (1.045) (0.323) (0.473) (0.348) Tax revenue over GDP 0.165∗∗ 0.183∗∗∗ 0.008 0.020 (0.067) (0.022) (0.036) (0.020) Percentage of population that uses the internet 0.037∗∗ 0.016∗∗ 0.045∗∗∗ 0.049∗∗∗ (0.017) (0.008) (0.012) (0.006) Share of the population age 65+ 0.689∗∗∗ 0.619∗∗∗ 0.006 − 0.006 (0.084) (0.033) (0.033) (0.044) Constant − 7.132 21.836∗∗∗ (7.182) (7.050) Test on the nonlinear subfunction of log GDP p.c. 3.704 2.226 p-value 0.000 0.013 Observations 1519 1518 1519 1518 Source: Analysis based on social protection expenditure data from the authors’ data set. Note: Standard errors are shown in parenthesis below the coefficients, clustered at the country level. ∗∗∗ indicates that the coefficient is significant at 1 percent level, ∗∗ at 5 percent level, ∗ at 10 percent level. All specifications include a correction for selective reporting using a cubic function of the propensity score (see supplementary online appendix table S1.2 for the probit regression results used to estimate the propensity scores). Figure 5 gives the SPEC when the analysis controls for all the covariates in table 4 set at the overall means, with and without country fixed effects. Without country effects, a rising SPEC still appears, though with a greatly attenuated slope relative to fig. 1. Now a mean SP share of around 9 percent of GDP among low-income countries is observed. Strikingly, when country effects are added, the income gradient reverses its sign over most of the range. In marked contrast to the “unconditional” SPEC in fig. 1, it is now found that the conditional mean share of GDP going to SP falls from 16 percent in the poorest countries to around 7 percent in the richest. While the study warns against attributing this dramatic change in the SPEC to any one covariate, it is noted that redoing fig. 5 only controlling for internet usage at its overall mean (along with country fixed effects) gives a similar picture (supplementary online appendix, fig. S1.5). With this one control for time-varying confounders, the SPEC is negatively sloped over most of the income range, falling from a mean share of 12 percent for the poorest country to 7 percent for the richest. Addressing Potential Selection Bias The rising SPEC could also be explained by a possible sample selection bias. Richer countries might have more complete SP spending data, which would impart an upward bias to the estimated income slope of the SPEC. The present study uses panel data on SP spending across countries and over time. It is a plausible assumption that the latent selection process increases the observed mean share of income devoted to SP spending conditional on mean income; all that is required for such positive selection is that higher SP/GDP values are more likely to be reported. However, that is not sufficient to explain why a rising SPEC is observed. Following the literature on estimating models with sample selection bias,33 let The study considers several competing methods for the lasso, such as the cv, minimizing BIC coefficients, adaptive lasso, and elastic lasso. The adaptive lasso technique was the most successful in terms of dimension reduction producing the specification with log GDP per capita, the share of the population older than 65, and the share of the population that uses the internet, which corresponds to the significant covariates shown in table 4. 33 See Vella’s (1998) survey. 538 Lokshin, Ravallion, and Torre Figure 5. Nonparametric SPECs Pooling Countries and Years with and without Country Fixed Effects and Controlling for All Time- Varying Covariates. 20 PLREG fixed effect PLREG no fixed effect Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 16 95% confidence interval SP expenditure, % of GDP 12 8 4 0 6.5 7 8 9 10 11 11.5 Log of GDP per capita (USD PPP) Source: Analysis based on social protection expenditure data from the authors’ data set. GDP data from WDI. Note: The graph plots SP spending as a percent of the GDP (vertical axis) against the log GDP per capita, in USD PPP prices (horizontal axis) for the pooled sample of 142 countries from 1995 to 2020. The (nonparametric) regression line controls for all covariates set their overall mean levels. g( p) denote an additive control function for selective reporting, which is considered a suitably smooth function of the probability of being selected into the sample, denoted p.34 Then, with positive selection (g ( p) > 0), the key issue in the present context is the sign and size of the correlation between the control function for selection and GDP per capita. A simple economic model illustrates how selective reporting arises in this context. Consider the costs and benefits of providing public spending data. (Without loss of generality, these can be taken to be the costs and benefits as perceived by public decision makers charged with this task.) It is plausible to assume that the marginal cost (MC) to a country’s government of collecting, processing, and disbursing such data rises with the amount of data made available. The marginal benefit (MB), on the other hand, can be expected to fall, if anything, as diminishing returns set in to collecting and processing extra data. It is also plausible that the MC of providing data tends to be higher for low-income countries than for high-income ones. This reflects the lower levels of state capacity generally in poorer countries, stemming from weaknesses in governance and ICT. The government chooses a level of data availability equating MC and MB, with the implication that low-income counties tend to make less data available as the MC function shifts upwards. 34 More precisely, p(Z ) = Prob(d = 1|Z ) is the propensity score for selection, where d = 1 if the country, year combination is observed while d = 0 otherwise, and Z is a vector of covariates for selection. The World Bank Economic Review 539 Then the control function for selection, g( p), can be expected to be positively correlated with GDP per capita. However, if one ignores the nonrandom selection, the control function is an omitted variable positively correlated with GDP per capita. In correcting for selective reporting, the study estimates the propensity scores (p-scores hereafter) using a probit regression for whether SP spending is data for a given country and year. In addition to (log) GDP per capita, the regressors are the log of the country population in that year and a variable measuring the Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 availability of SP spending data for that country in prior years.35 By including population size the study’s reasoning is that public spending data have properties of a public good, such that lower average costs in more-populous countries entail that the data are more likely to be available. The variable for past SP data is motivated by the assumption that past experience in providing SP spending data makes it less costly to currently provide such data. A cubic function of the p-scores is then included in the regressions for the SPEC as control function.36 All three regressors in the probit for the availability of SP spending data, namely log GDP per capita, log population, and the dummy variable for whether SP spending data had been available in the data set, have the expected signs and are significant at the 5 percent level or better; the pseudo-R2 was 0.423 (table S1.2.) The income effect remains strong with the correction for bias due to selective reporting – the coefficient on log GDP per capita is 3.064 in the OLS regression and 1.361 in the FE regression, both statistically significant at the 1 percent level. So, the selective reporting bias could not explain the rising SPEC.37 5. Testing the Hypotheses Using Data on SP Responses to the Pandemic “Descriptive Statistics on SP Spending” in section 3 showed that the SPECs have not changed over time in a way that is suggestive of a shift in development priorities in poor and middle-income countries such that more of their national income is devoted to SP. Yes, there has been an increase in the SP share of GDP over time among initially poor countries, but this has been driven by the fact that they are no longer quite so poor. However, (as noted above) there may well be a degree of stickiness over time in the SPECs. (Line ministries are known to be resistant to cuts in their budgets.) Possibly public spending decisions adjust slowly to the new priorities. The COVID shock induced SP responses in varying degrees across the world, as documented by Gentilini et al. (2021). This shock-induced change in policies may better reveal the true change in the underlying policy priorities (even prior to the pandemic), in that the pandemic gave policy makers a new opportunity to implement the policies they truly want – a break from the past. Studying the pandemic responses might also provide a new perspective on what country-level factors influence the success of pro-SP policy reforms. It may well be that the country characteristics relevant to SP spending in normal times differ from those that matter most in responding to a large shock or in policy reforms. For example, ICT infrastructure is likely to matter to the scope for scaling up existing programs, while general government effectiveness may matter more to the capabilities for introducing new programs, which can be administratively demanding. 35 More precisely, this variable measures the share of previous country-year observations (from 1995 and up to the year previous to that of each observation) where SP spending data is not missing. When data are not missing in every year since 1995 the variable takes a value of 1, and when data is always missing the variable takes a value of 0. The variable takes intermediate values whenever information in some years is missing and nonmissing in others. 36 For a more general treatment of these methods of correcting for sample selection bias, see Das, Newey, and Vella (2003). 37 Note that only the last of the regressions in table 3 corrects for sample selection bias, since the study treats this as a distinct hypothesis rather than in combination with other hypotheses. The encompassing regression in table 4 corrects for sample selection as a member of the set of joint hypotheses. 540 Lokshin, Ravallion, and Torre The COVID-19 pandemic is not the first economic crisis in recent times. There is evidence that, at least in OECD countries, the Global Financial Crisis (GFC) of 2008 increased social spending (to a large degree through automatic stabilizers) (e.g., ILO and World Bank 2012). The study tried replicating its analysis using data on the SP policy responses undertaken by governments around the world in the aftermath of the financial crisis (fig. S1.6 in the supplementary online appendix). The SPEC of stimulus spending in the GFC is rising with per capita income, and government effectiveness attenuates the SPEC’s slope as it does Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 for the pandemic stimulus SPEC. However, the data on SP stimulus in the GFC are available only for 67 countries, which is almost three times fewer than the sample size of the dataset that the study assembled for the COVID-19 SP stimulus expenditures; the confidence intervals of the SPEC estimations for the GFC are too large to draw reliable conclusions.38 The study finds that poorer countries devoted a smaller share of their GDP to social protection in response to the pandemic. The bold line in fig. 6 plots the relationship.39 (The dashed line is explained below.) The pattern is similar to fig. 1, although the mean shares of GDP are lower when the analysis focuses on the pandemic response alone. Similarly to fig. 1, the analysis still sees a marked tendency for the share of GDP devoted to the SP response to rise with GDP per capita. The factor of three or more in the ratio of mean SP share for high-income countries over that for low-income countries is also evident in fig. 6.40 The data on SP spending in response to the pandemic form a single cross-sectional data set, as distinct from the longitudinal (panel) data that were assembled for looking at SP spending prior to the pandemic. So, country fixed effects cannot be included. However, as a check for bias associated with latent country characteristics, the analysis includes the estimated country effect from the pre-pandemic panel data re- gressions as an extra control in the encompassing model.41 Notice that, on a priori grounds, the sign on this latent effect could go either way. On the one hand, countries with greater prior experience with SP interventions may have seen lower costs of implementing an extra SP effort during the pandemic. But, on the other hand, they would presumably have less need for extra SP effort during the pandemic. Similarly to its analysis of the pre-pandemic data, the study builds up its tests using regression controls to augment fig. 6. Table 5 provides the results of testing the hypotheses one at a-time. Relative prices for education and health services do not shift the SPEC much. Mixed support is found for the Accountability Hypothesis. In the linear regression, the VA and Polyarchy indices have positive coefficients, significant at the 5 percent level, but these effects are not robust to allow for nonlinearity using the PLREG estimator. The GDP effect remains strong, however, so it cannot be concluded that the Accountability Hypothesis explains that effect. There is stronger support for the Governance 2 Hypothesis in the SP response to the pandemic. Strik- ingly, GE knocks out GDP per capita, leaving it with a small effect not significantly different from zero. The GE indicator has a positive and significant effect in both the linear and nonlinear SPECs. 38 For this analysis, the study relies on the data included in the ILO/World Bank Inventory of policy responses to the global financial and economic crisis of 2008 (ILO and World Bank 2012). This dataset includes the expenditure on policies implemented in response to the Global Financial Crisis over a period of two years (from mid 2008 to mid 2010). It mirrors, in structure, the dataset the study uses for the COVID-19 analysis, which relies on a similar inventory of policy responses and the associated expenditure over a period of one year and a half. 39 This is an updated version of a similar graph found in Demirgüç-Kunt et al. (2022), though the latter paper used a linear functional form in representing the line of best fit. 40 Granted, there are some notable differences; for example, the United States is well below the regression line in fig. 1, but well above it in fig. 6. However, taken overall, the study does not find that the pandemic-induced policy shock generated a cross-country SPEC that looks very different to what was found prior to the pandemic. 41 Since this is a generated regressor, standard errors are bootstrapped. The analysis uses the country effects estimated in the OLS specification of pre-pandemic data (third column in table 4). The World Bank Economic Review 541 Figure 6. Social Protection Engel Curve during the pandemic. 12 Lowess regression, no controls USA PLREG with control for government effectiveness SP response budget as a proportion of GDP (%) 95% confidence interval Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 TLS 9 SRB JPN SGP CAN AUS 6 NZL THA GEO SYC IRL AUT CHN LVA ISR FIN LTU UKR HRV GBR LUX 3 KHM MNG ITA CHE PHL ARG PAK IND ISL MLI CZE SOM COM DOM NOR LBR CPV MMR MEX ARM URY BFA MAR SAU COD TTO QAT 0 STP LAO GNQ 6.5 7 8 9 10 11 12 Log of GDP per capita (2019 PPP) Source: Analysis based on social protection pandemic response data from Demirgüç-Kunt et al. (2022). Government effectiveness data from World Bank (2022a). GDP data from WDI. Note: The graph plots, for 154 countries with the required data, the social protection response budget to the COVID-19 pandemic, as a percent of the 2019 GDP (vertical axis) against log GDP per capita for 2019, in USD PPP prices (horizontal axis). The dashed line gives the nonparametric SPEC when one controls for the Government Effectiveness indicator (set at its global mean value). Support is also found for the Technology Hypothesis, with mobile phone usage having a significant effect at the 5 percent level; the GDP effect is attenuated somewhat – when mobile phones subscriptions are included as control – but remains strong. In contrast to the pre-pandemic SP spending data, no significant effect of internet usage in the pandemic responses is found. There is support for the Ageing Hypothesis in the linear specification, with the population share over 65 having a significant effect at the 5 percent level, but this is not robust to allowing for nonlinearity in the SPEC. Table 5 provides the test of whether it is the COVID impact that explains the income effect. COVID mortality does not appear to be the reason for the income effect in fig. 6; the study does not see higher SP spending by richer countries because they were hit harder by the pandemic.42 42 The reporting and recording of the causes of death could be less precise or biased in the poor countries, especially during times when COVID cases were peaking, and health systems were most strained. But if the cases of COVID death were underreported in poor countries, this would reduce the effect of this control even more, reinforcing the current conclusion. 542 Lokshin, Ravallion, and Torre Table 5. Separate Tests Using SP Responses to the Pandemic. Estimation Log GDP per capita Test on Control variable Number of Hypothesis method Coefficient Std. error f (lnGDPit ) Coefficient Std. error observations Baseline OLS 0.840∗∗∗ 0.148 154 Relative prices 1 Health/Food price ratio OLS 0.820∗∗∗ 0.182 0.783 0.794 146 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 PLREG 2.035†† 0.030 1.233 146 Relative prices 2 Education/Food price ratio OLS 0.798∗∗∗ 0.183 0.584 0.505 146 PLREG 1.863†† − 0.032 0.822 146 Accountability 1 Polyarchy index OLS 0.701∗∗∗ 0.149 0.061∗∗ 0.029 141 PLREG 0.345 0.059 0.042 141 Accountability 2 Political competition index OLS 0.726∗∗∗ 0.149 0.099 0.062 141 PLREG 0.447 0.075 0.090 141 Accountability 3 Voice and Accountability OLS 0.632∗∗∗ 0.181 0.449∗∗ 0.227 154 †† Index PLREG 0.987 0.359 0.294 154 Governance: Govt. effectiveness Government effectiveness OLS − 0.217 0.278 1.465∗∗∗ 0.333 154 Index PLREG 0.129 1.617∗∗∗ 0.454 154 Governance: Domestic resources Tax revenue over GDP OLS 0.713∗∗∗ 0.175 0.038 0.030 125 PLREG 0.267 0.026 0.039 125 Technology 1 Mobile phone suscriptions OLS 0.606∗∗∗ 0.186 0.013∗∗ 0.006 150 per 100 inhabitants PLREG 1.082††† 0.014∗∗ 0.007 150 Technology 2 Percentage of population OLS 0.934∗∗∗ 0.395 − 0.004 0.017 151 that uses internet PLREG 0.534 − 0.001 0.020 151 Aging Share of the population OLS 0.492∗∗∗ 0.213 0.105∗∗ 0.045 147 age 65+ PLREG 0.594 † 0.089 0.060 147 Covid-19 hypothesis Total COVID19 death OLS 0.931 0.166 − 0.245 0.197 153 per 1,000,000 PLREG 2.643††† − 0.029 0.265 153 Source: Analysis based on social protection pandemic response data from Demirgüç-Kunt, Lokshin, and Torre (2022). Note: Standard errors are clustered at the country level. ∗∗∗ indicates that the coefficient is significant at 1 percent level, ∗∗ at 5 percent level, ∗ at 10 percent level. ††† indicates the test p-value <0.001, †† p-value <0.01, and † p-value < 0.1. So the government effectiveness index is the only covariate that can, on its own, robustly account for the income effect of the SP share in pandemic responses. And the study finds that the GE effect dominates the income effect. The dashed line in fig. 6 plots the relationship with GDP but now with the control for GE (set at the overall mean) using the PLREG estimator to ensure flexibility in representing the income effect on the SPEC. Once GE is controlled for, the share of GDP devoted to SP during the pandemic declines as GDP per capita rises, although statistically, one cannot reject the null hypothesis of a 0 slope (table 5). The quantitative effect of GE differences is sizeable. Consider the 20 countries with the lowest GDP per capita in the data set. If their GE index was the same as the overall mean index, then their share of GDP devoted to SP would have risen from 0.86 to 2.05 percent of GDP. To this point, each hypothesis has been tested one at a time. Table 6 combines the controls, and also provides the PLREG estimates of equation (2), allowing a flexible representation of the GDP effect. The The World Bank Economic Review 543 Table 6. Encompassing Regressions for SP Responses to the Pandemic. Baseline specification With imputed fixed effects OLS PLREG OLS PLREG Log GDP per capita GDP − 0.024 0.116 (0.458) (0.598) Voice and accountability − 0.319 − 0.370 − 0.492 − 0.419 Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 (0.313) (0.363) (0.346) (0.385) Government effectiveness 1.544∗∗∗ 1.957∗∗∗ 1.591∗∗∗ 1.750∗∗∗ (0.480) (0.599) (0.532) (0.637) Percentage of population that use internet − 0.017 − 0.011 − 0.020 − 0.021 (0.018) (0.021) (0.022) (0.029) Share of population over 65 years 0.096∗ 0.088 0.173∗∗∗ 0.172∗∗ (0.052) (0.064) (0.064) (0.081) Covid-19 deaths per million/1000 − 0.121 − 0.015 − 0.108 − 0.028 (0.222) (0.292) (0.221) (0.275) Imputed country effect − 0.089∗ − 0.116∗ (0.048) (0.060) Constant 2.591 0.712 (3.624) (4.649) Test on log GDP subfunction in PLREG 0.323α 0.432α 0.375β 0.333β R2 0.285 0.329 Number of countries 147 120 Source: Analysis based on social protection pandemic response data from Demirgüç-Kunt, Lokshin, and Torre (2022). Note: ∗∗∗ indicates that the coefficient is significant at 1 percent level, ∗∗ at 5 percent level, ∗ at 10 percent level. Standard errors are shown in parenthesis below the coefficients, clustered at the country level. Bootstrapped standard errors for the specifications using imputed fixed effects. α Value of the test that Log GDP effect is significant. β Probability that test value is different from zero. government effectiveness index not only knocks out the GDP effect, it is also the (statistically) strongest control variable in explaining the SP shares during the pandemic. There is also support for the Age- ing Hypothesis in the encompassing model, once latent country effects are included in the SPEC (based on the retained country effects from the corresponding regressions in section 4), which have a negative sign. It has been seen that SP spending in response to the pandemic showed a very similar relationship with national income to pre-pandemic SP spending, though the covariates are somewhat different, with a more important role played by government effectiveness and (much) less by ICT. However, SP was only one component of the policy response to the pandemic; extra spending also took place in health and infrastructure, for example. If the constraints on ensuring effective SP in poor countries are not as severe for other (non-SP) pandemic policies, then the governments of poorer countries may have been drawn to substitute toward other types of spending in their response to the pandemic. That conjecture is not consistent with what the data show. Figure 7(a) provides the Engel curve for total stimulus spending while fig. 7(b) provides it for non-SP stimulus spending. It can be seen that the pattern is similar during the pandemic (comparing to fig. 1) although the (unconditional) pandemic SPEC is flatter until the upper-middle incomes area is reached. When government effectiveness is controlled for, the share of GDP devoted to both total stimulus spending and its non-SP components shows a U-shaped relationship, with relatively high shares implied for poor countries. It can be inferred that the constraints stemming from weak governance applied not only to SP spending, but also impacted non-SP spending responses to the pandemic. 544 Lokshin, Ravallion, and Torre Figure 7. Total Stimulus Spending and Non-SP Spending during the Pandemic. (a) Total Stimulus Spending. Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Source: Analysis based on social protection pandemic response data from Demirgüç-Kunt, Lokshin, and Torre (2022). Government effectiveness data from World Bank (2022a). GDP data from WDI. Note: The graph plots, for 154 countries with the required data, the total response budget to the COVID-19 pandemic, as a percent of the 2019 GDP (vertical axis) against log GDP per capita for 2019, in USD PPP prices (horizontal axis). The dashed line gives the nonparametric SPEC when one controls for the Government Effectiveness indicator (set at its global mean value). (b) Non-SP spending. Source: Analysis based on social protection pandemic response data from Demirgüç-Kunt, Lokshin, and Torre (2022). Government effectiveness data from World Bank (2022a). GDP data from WDI. Note: The graph plots, for 154 countries with the required data, the nonsocial protection response budget to the COVID-19 pandemic, as a percent of the 2019 GDP (vertical axis) against log GDP per capita for 2019, in USD PPP prices (horizontal axis). The dashed line gives the nonparametric SPEC when one controls for the Government Effectiveness indicator (set at its global mean value). The World Bank Economic Review 545 6. Conclusions Poor countries devote a much lower share of their national income to social protection than rich countries do; in other words, a rising Engel curve for social protection across countries can be seen. This has led some observers to argue that social protection spending is a luxury good. However, explaining a phenomenon by differences in preferences hardly leads to meaningful policy conclusions. This paper has proposed a number of alternative hypotheses that are suggestive of potentially confounding time-varying variables, Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 which can be interpreted as differences in the costs incurred by governments in supplying social protection. There are undoubtedly other covariates that this study has not adequately accounted for.43 However, it is not necessary to account for all possible confounders; along with country fixed effects, the confounders identified in this paper not only reduce the slope of the SP Engel curve, they also change its sign, suggesting that the conditional mean share falls over a wide range of incomes. Nor does the fact that richer countries have been observed to devote a higher share of their GDP to social protection during the pandemic imply that social protection is a luxury good. Here again, this paper finds that omitted variables are confounding the “preferences” interpretation. When it comes to implementing SP policy responses to the pandemic, or any other big shock, the effectiveness of government in delivering public services more generally can be expected to be a decisive factor. Scaling up existing SP programs will no doubt play a role, but responses to a shock will often require rapid resource mobilization and the ability to design and implement new policies (with new target beneficiaries, such as those whose employment is at risk) – all of which will be easier with greater (pre-pandemic) capabilities for effective public service delivery. Indeed, this paper has shown that if one controls only for government effectiveness (set at the global mean), then the share of GDP devoted to social protection during the pandemic is essentially no different comparing rich countries with poor ones. These results suggest that the upward-sloping SPEC reflects weaknesses in governmental effectiveness generally – weaknesses that are highly correlated with average income but still have a degree of independent variation. The paper’s results point to the importance of exploring further how broader efforts to improve gover- nance and access to technology in developing countries may help attain better social protection. Without success in such efforts, poor countries may be caught in a vicious cycle whereby weakness in these corre- lates of low income inhibits effective social protection, which helps maintain poverty. Appendix A1. Data Sources on SP Spending This paper’s dataset on social protection expenditures covers 142 countries over the period 1995–2020. The study defines social protection expenditure as the public expenditure covering social assistance, social insurance (including pensions), and active labor market policies. The dataset contains 2,481 country-year observations coming from the sources listed in table A1.1. For countries in the European Union, including the United Kingdom, the analysis uses data from Eu- rostat’s SP expenditure database (table code: spr_exp_gdp). To be consistent with the definition of SP, the study excludes the expenditure under ESSPROS function “sickness and healthcare.” These data cover the whole period 1996–2020 for EU-15 countries, and for most of the period for the remaining EU member states. In total, Eurostat data is used for 587 country-year observations. For countries in Europe and Central Asia not covered by Eurostat data, the study uses the World Bank’s Social Protection Expenditure and Evaluation Database (SPEED), which covers 27 countries in that region (including Turkey) from 2000 to 2020. In total, SPEED is used for 380 country-year observations. 43 For example, one could consider testing the relationship between trade openness and social spending as in, e.g., Avelino et al. (2005). 546 Lokshin, Ravallion, and Torre Table A1.1. Sources for SP Expenditure. Country-year Number of Source observations countries Period Regions Eurostat 587 28 1996–2019 European Union + United Kingdom SPEED 380 27 2000–2020 Europe and Central Asia + Turkey OECD 73 11 2000–2019 Non-European OECD countries Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 ECLAC 443 32 1995–2020 Latin America and the Caribbean GFS (IMF) 782 94 1995–2020 Africa, Middle East, South Asia, East Asia, and the Pacific + countries not included in above sources World Bank PERs 179 26 1995–2020 Africa, Middle East, South Asia, East Asia, and the Pacific World Bank 37 3 2004–2018 Benin, Mali, and Tunisia BOOST Total 2,481 142 1995–2020 For OECD countries outside Europe, the study uses data from OECD’s Social Expenditure database. This database provides information for 11 countries in the period 2000 to 2019. In total OECD data are used for 73 country-year observations. For countries in Latin America and the Caribbean not included in the OECD’s Social Expenditure database, the study uses data from the Economic Development Division of the Economic Commission for Latin America and the Caribbean (ECLAC). These data correspond to government expenditure classified by function (COFOG). When available, the analysis use the values corresponding to the SP expenditure by the general government; otherwise, it uses the values corresponding to the SP expenditure by the central government. This source provides information for 32 countries in the period 1995–2020. In total, ECLAC data is used for 443 country-year observations. For countries in the remaining regions of the world – Africa, Middle East, South Asia, and East Asia and the Pacific – or that are not covered by any of the above databases, two sources are relied on: the IMF’s Government Finance Statistics and the World Bank’s Public Expenditure Reviews. The IMF’s Government Finance Statistics provide data on government expenditure classified by func- tion (COFOG). When available, the analysis uses the values corresponding to the SP expenditure by the general government; otherwise, it uses the values corresponding to the SP expenditure by the central gov- ernment including social security funds. This source provides information for 94 countries in the period 1995–2020. In total, GFS data for 782 country-year observations are used. The World Bank Public Expenditure Reviews (PERs) are a series of non-periodical reports analyzing the public expenditure of World Bank client countries. These reports are occasional and do not follow a fixed outline (unlike the IMF’s Article IV reviews). Most of these reports contain information on SP expenditure. The study uses information coming from different PERs for 26 countries. In total, it uses PER data for 179 country-year observations. Lastly, for Benin, Mali, and Tunisia the data come from the World Bank’s BOOST Open Budgets Portal. This portal presents detailed budget information for these countries, which makes it possible to calculate public expenditure on SP. In total, BOOST is used for 37 country-year observations. References Acemoglu, D., S. Johnson,, and J. Robinson, 2005. “Institutions as a Fundamental Cause of Long-Run Growth.” In Handbook of Economic Growth,Vol. IA.Edited by Aghion, P, and S. N. Durlauf, 386–472. Amsterdam: Elsevier. Alesina, A., and F. Passarelli, 2019. “Loss Aversion in Politics.” American Journal of Political Science 63(4): 936–947. Alesina, A., and D. Rodrik, 1994. “Distributive Politics and Economic Growth.” Quarterly Journal of Economics 108: 465–90. The World Bank Economic Review 547 Angrist, J., and J.-.S. Pischke, 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton: Princeton University Press. Auteri, M., and M. Constantini, 2004. “Is Social Protection a Necessity or a Luxury Good? New Multivariate Coin- tegration Panel Data Results.” Applied Economics 36: 1887–98. Avelino, G., D.S. Brown, and W. Hunter, 2005. “The Effects of Capital Mobility, Trade Openness, and Democracy on Social Spending in Latin America, 1980–1999.” American Journal of Political Science 49(3): 625–41. Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Belloni, A., V. Chernozhukov,, and C. Hansen, 2014. “High-Dimensional Methods and Inference on Structural and Treatment Effects.” Journal of Economic Perspectives 28(2): 29–50. Besley, T., and T. Persson, 2011. Pillars of Prosperity: The Political Economy of Development Clusters. Princeton: Princeton University Press. Bird, R., 1971. “Wagner’s Law of Expanding State Activity.” Public Finance 26: 1–26. Brückner, M., A. Chong, and M. Gradstein, 2012. “Estimating the Permanent Income Elasticity of Government Ex- penditures: Evidence on Wagner’s Law Based on Oil Price Shocks.” Journal of Public Economics 96(11–12): 1025– 1035. Center for Systemic Peace, 2020, “ Polity5: Political Regime Characteristics and Transitions 1800–2018.” Data set available at. https://www.systemicpeace.org/inscrdata.html Clemente, J., C. Marcuello,, and A. Montañes, 2012. “Government Social Spending and GDP: Has there been a Change in Social Policy?” Applied Economics 44(22): 2895–905. Coppedge, M., J. Gerring, C.H. Knutsen, S.I. Lindberg, J. Teorell, K.L. Marquardt, J. Medzihorsky, D. Pemstein, L. Gastaldi, S. Grahn, J. Pernes, O. Rydén, J. Römer, E. Tzelgov, Y. Wang, and S. Wilson, 2023. “V-Dem Methodology v13” Varieties of Democracy (V-Dem) Project. Dataset is available at https://www.v-dem.net/data/the-v-dem-dat aset/. Cornelisse, P., and K. Goudswaard, 2002. “On the Convergence of Social Protection Systems in the European Union.” International Social Security Review 55(3): 3–17. Das, M., W. Newey, and F. Vella, 2003. “Nonparametric Estimation of Sample Selection Models.” Review of Economic Studies 70(1): 33–58. Demirgüç-Kunt, A., M. Lokshin,, and I. Torre, 2022. “Protect Incomes or Protect Jobs? The Role of Social Policies in Post-Pandemic Recovery.” Policy Research Working Paper #10166. The World Bank. Washington, DC, USA. Dincecco, M., 2009. “Fiscal Centralization. Limited Government, and Public Revenues in Europe, 1650–1913.” Jour- nal of Economic History 69(1): 48–103. Gentilini, U., M. Almenfi, I. Orton, and P. Dale, 2021. “Social Protection and Jobs Responses to Covid-19.” Living Paper. World Bank. Washington, DC, USA. Gómez, R., and P. Hernández de Cos, 2008. “The Importance of Being Mature: The Effect of Demographic Maturation on Global per Capita GDP.” Journal of Population Economics 21: 589–608. Han, K., 2021. “Political Budgetary Cycles in Autocratic Redistribution.” Comparative Political Studies 55: 727–56. International Labor Organization and World Bank. 2012. Inventory of Policy Responses to the Financial and Eco- nomic Crisis. Geneva, Switzerland, and Washington, DC, USA. International Monetary Fund (IMF), 2021, “Fiscal Monitor Database of Country Fiscal Measures in Response to the COVID-19 Pandemic.” Fiscal Affairs Department, International Monetary Fund: Washington, DC, USA. International Comparison Program (ICP) database. https://www.worldbank.org/en/programs/icp. Jack, W., A. Ray, and T. Suri, 2013. “Transaction Networks: Evidence from Mobile Money in Kenya.” American Economic Review 103(3): 356–61. Jolliffe, D., and E. Prydz, 2016. “Estimating International Poverty Lines from Comparable National Thresholds.” Journal of Economic Inequality 14(2): 185–98. Kristov, L., P. Lindert,, and R. McClelland, 1992. “Pressure Groups and Redistribution.” Journal of Public Economics 48(2): 135–63. Leser, C.E.V., 1963. “Forms of Engel Functions.” Econometrica 31: 694–703. Lindert, P., 1994. “The Rise of Social Spending, 1880–1930.” Explorations in Economic History 31(1): 1–37. Lokshin, M., 2006. “Difference-Based Semiparametric Estimation of Partial Linear Regression Models.” Stata Journal: Promoting Communications on Statistics and Stata 3: 377–383. Meltzer, A., and S. Richard, 1981. “A Rational Theory of the Size of Government.” Journal of Political Economy 89(5): 914–27. 548 Lokshin, Ravallion, and Torre Milanovic, B., 2016. Global Inequality: A New Approach for the Age of Globalization. Cambridge, MA: Harvard University Press. Mulligan, C., R. Gil, and X. Sala-i-Martin, 2004. “Do Democracies Have Different Public Policies than Nondemoc- racies?” Journal of Economic Perspectives 18(1): 51–74. Peracchi, F., 2001, “Patterns of Social Spending in Western Europe.” Tor Vergata University. Rome, Italy. Preston, S., 1975. “The Changing Relationship Between Mortality and Level of Economic Development.” Population Downloaded from https://academic.oup.com/wber/article/37/4/519/7238664 by University of Oxford user on 12 December 2023 Studies 29: 231–48. Ravallion, M., 2016a, The Economics of Poverty: History, Measurement and Policy. New York: Oxford University Press. ———, 2016b, “Toward Better Global Poverty Measures.” Journal of Economic Inequality 14: 227–248. ———, 2017, Interventions against Poverty in Poor Places.” 20th Annual WIDER Lecture. World Institute of Devel- opment Economics Research: Helsinki, Finland. ———, 2021, What Might Explain Today’s Conflicting Narratives on Global Inequality?” In Inequality in the Devel- oping World. Edited by Gradin, C., M. Leibbrandt, and F. Tarp, 17–48. Oxford: Oxford University Press. Ravallion, M., and S. Chen, 2019. “Global Poverty Measurement when Relative Income Matters.” Journal of Public Economics 177: 1–13. Ravallion, M., S. Chen, and P. Sangraula, 2009. “Dollar a Day Revisited.” World Bank Economic Review 23(2):163– 84. Ritchie, H., and M. Roser, 2019. “Age Structure.” Our World in Data. Oxford University. https://Age Structure - Our World in Data. Shelton, C., 2007. “The Size and Composition of Government Expenditure.” Journal of Public Economics 91: 2230– 60. Vella, F., 1998. “Estimating Models with Sample Selection Bias: A Survey.” Journal of Human Resources 33(1): 127– 169. Williamson, J., 1990, “What Washington Means by Policy Reform,” In Latin American Readjustment: How Much has Happened. Edited by Williamson, J., 7–20. Washington, DC: Institute for International Economics. Wooldridge, J.. 2001. Econometric Analysis of Cross-Section and Panel Data. Cambridge, MA: MIT Press. Working, H., 1943, “Statistical Laws of Family Expenditures,” Journal of the American Statistical Association 38: 43–56. World Bank, 1990, World Development Report: Poverty. New York: Oxford University Press. ———, 2006. World Development Report: Equity and Development. Washington DC: World Bank. ———, 2022a, “Worldwide Governance Indicators.” World Bank. Washington, DC, USA. ———, 2022b. “G2Px: Digitizing Government-to-Person Payments.” Washington, DC: World Bank. ———, 2023. Worldwide Governance Indicators. http://info.worldbank.org/governance/wgi/. World Development Indicators (WDI) database. https://wdi.worldbank.org/. Yatchew, A., 1998. “Nonparametric Regression Techniques in Economics.” Journal of Economic Literature 36: 669– 721. Yu, X., 2016, “Social Protection Tools for the 21st Century.” Voices. World Bank Blog. https://blogs.worldbank.org/ voices/social- protection- tools- 21st- century.