WPS8593 Policy Research Working Paper 8593 Background Paper to the 2019 World Development Report Methodology for a World Bank Human Capital Index Aart Kraay Development Economics Development Research Group September 2018 Policy Research Working Paper 8593 Abstract This paper describes the methodology for a new World Bank of complete education and full health, and ranges from 0 Human Capital Index (HCI). The HCI combines indica- to 1. A value of x on the HCI indicates that a child born tors of health and education into a measure of the human today can expect to be only x ×100 percent as productive capital that a child born today can expect to obtain by her as a future worker as she would be if she enjoyed complete 18th birthday, given the risks of poor education and health education and full health. The methodology of the HCI that prevail in the country where she lives. The HCI is is anchored in the extensive literature on development measured in units of productivity relative to a benchmark accounting. This paper—prepared as a background paper to the World Bank’s World Development Report 2019: The Changing Nature of Work—is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The author may be contacted at akraay@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team     Methodology for a World Bank Human Capital Index      Aart Kraay1                    JEL Codes: I1, I2, O1, O4   Keywords:  human capital, health, education, development accounting                                                                 1  World Bank Development Research Group, akraay@worldbank.org.  This paper was prepared as a background  paper for the World Development Report 2019 and for the World Bank’s Human Capital Project.  It has benefited  from extensive discussions with Roberta Gatti, Simeon Djankov and David Weil (Brown).  Particular thanks to  Rachel Glennerster (DFID), Bill Maloney, Mamta Murthi and Martin Raiser for peer review comments; to Chika  Hayashi (UNICEF) and Espen Prydz for guidance on stunting data; to Noam Angrist (Oxford), Harry Patrinos and  Syedah Aroob Iqbal for the harmonized test score data; to Deon Filmer and Halsey Rogers for extensive discussions  on converting test scores into learning‐adjusted school years; to Husein Abdul‐Hamid, Anuja Singh (UNESCO) and  Said Ould Ahmedou Voffal (UNESCO) for help with enrollment data; to Patrick Eozenou and Adam Wagstaff for  help with DHS data; and to Krycia Cowling (IHME), Nicola Dehnen and Ritika D’Souza for tireless research  assistance.  Valuable comments were provided by Sudhir Anand (Oxford), George Alleyne (PAHO), Ciro Avitabile,  Francesco Caselli (LSE), Matthew Collins, Shanta Devarajan, Patrick Eozenou, Tim Evans, Jed Friedman, Emanuela  Galasso, Michael Kremer (Harvard), Lant Pritchett (Harvard), Federico Rossi (Johns Hopkins), Michal Rutkowski,  Jaime Saavedra, Adam Wagstaff, and Pablo Zoido‐Lobatón (IDB).  This paper has also benefitted from the  discussion at two workshops on measuring the contribution of health to human capital held at the World Bank on  March 1, 2018 and May 14, 2018, and a Bank‐wide review meeting held June 11, 2018.  The data used in this paper  have benefitted from an extensive consultation process organized by the office of the World Bank Chief Economist  for Human Development, which resulted in many expansions and refinements to the school enrollment and  stunting data used in the HCI.  The HCI will be published in the 2019 World Development Report and  accompanying special report on human capital.  The views expressed here are the author’s, and do not reflect those  of the World Bank, its Executive Directors, or the countries they represent.  1.  Introduction    Effective investments in human capital are central to development, delivering substantial  economic benefits in the long term.  However, the benefits of these  investments often take time to  materialize and are not always very visible to voters.   This is one reason why policymakers may not  sufficiently prioritize programs to support human capital formation.  At the 2017 Annual Meetings,  World Bank management called for a Human Capital Project (HCP) to address this incentive problem  through a program of advocacy and analytical work intended to raise awareness of the importance of  human capital and to increase demand for interventions to build human capital in client countries.    The advocacy component of the HCP features a Human Capital Index (HCI) that measures the  human capital that a child born today can expect to attain by age 18, given the risks to poor health and  poor education that prevail in the country where she lives.  The HCI is designed to highlight how  investments that improve health and education outcomes today will affect the productivity of future  generations of workers.  The HCI measures current education and health outcomes since they can be  influenced by current policy interventions to improve the quantity and quality of education, and health.      The main text of this paper provides a nontechnical description of the components of the HCI  (Section 2) and how they are combined into an aggregate index (Section 3).  This is followed by a  description of the index and its interpretation (Section 4).  Section 5 discusses how the index can be  linked to aggregate per capita income differences and growth, and Section 6 concludes.  A lengthy  technical appendix provides details on index methodology and data, as well as citations to the relevant  literature.    2.  Components of the Human Capital Index    Imagine the trajectory from birth to adulthood of a child born today.  In the poorest countries in  the world, there is a significant risk that the child does not survive to her fifth birthday.  Even if she does  reach school age, there is a further risk that she does not start school, let alone complete the full cycle  of 14 years of school from pre‐school to Grade 12 that is the norm in rich countries.  The time she does  spend in school may translate unevenly into learning, depending on the quality of teachers and schools  she experiences.  When she reaches age 18, she carries with her lasting effects of poor health and  nutrition in childhood that limit her physical and cognitive abilities as an adult.     The goal of the HCI is to quantitatively illustrate the key stages in this trajectory and their  consequences for the productivity of the next generation of workers, with these three components:  2    Component 1:  Survival.  This component of the index reflects the unfortunate reality that not all  children born today will survive until the age when the process of human capital accumulation through  formal education begins.  It is measured using under‐5 mortality rates taken from the UN Child Mortality  Estimates (Figure 1), with survival to age 5 as the complement of the under‐5 mortality rate.  Data on  under‐5 mortality are available for 198 countries, and much of the variation across countries in child  mortality rates reflects differences in mortality in the first year of life.  Component 2: Expected Learning‐Adjusted Years of School.  This component of the index combines  information on the quantity and quality of education.  The quantity of education is measured as the  number of years of  school a child can expect to obtain by age 18 given the prevailing pattern of  enrolment rates. It is calculated as the sum of age‐specific enrollment rates between ages 4 and 17.   Age‐specific enrollment rates are approximated using school enrollment rates at different levels:  pre‐ primary enrollment rates approximate the age‐specific enrollment rates for 4 and 5 year‐olds; the  primary rate approximates for 6‐11 year‐olds; the lower‐secondary rate approximates for 12‐14 year‐ olds; and the upper‐secondary rate approximates for 15‐17 year‐olds.  Data to construct this measure is  available for 194 countries (Figure 2).  The quality of education reflects new work at the World Bank to  harmonize test scores from major international student achievement testing programs (Figure 2).  The  database covers over 160 countries.  These are combined into a measure of expected learning‐adjusted  years of school, using the conversion metric proposed in the 2018 World Development Report (Figure  3).   Component 3:  Health  There is no single broadly‐accepted, directly‐measured, and widely‐available  metric of health that is analogous to years of school as a standard metric of educational attainment.  In  the absence of such a measure, two proxies for the overall health environment are used to populate this  component of the index:  (i) adult survival rates, defined as the fraction of 15 year‐olds that survive until  age 60, and (ii) the rate of stunting for children under age 5 (Figure 4).  Adult survival rates are  calculated by the UN Population Division for 197 countries.  In the context of the HCI they are used as a  proxy for the range of non‐fatal health outcomes that a child born today would experience as an adult if  current conditions prevail into the future.  Stunting serves as an indicator for the pre‐natal, infant and  early childhood health environment, summarizing the risks to good health that children born today are  likely to experience in their early years – with important consequences for health and well‐being in  adulthood.  Data on the prevalence of stunting is reported in the UNICEF‐WHO‐World Bank Joint  Malnutrition Estimates.  This dataset contains 132 countries with at least one estimate of stunting in the  3    past 10 years.   The considerations leading to the choice of these two proxy measures for the overall  health environment are detailed in Appendix A3.  3.  Aggregating the Components into a Human Capital Index     The health and education components of human capital all have intrinsic value that is  undeniably important but difficult to quantify.  This in turn makes it challenging to combine the different  components into a single index.   One solution that permits aggregation is to interpret each component  in terms of its contribution to worker productivity, relative to a benchmark corresponding to complete  education and full health.      In the case of survival, the relative productivity interpretation is very stark, since children who  do not survive childhood never become productive adults.  As a result, the expected productivity as a  future worker of a child born today is reduced by a factor equal to the survival rate, relative to the  benchmark where all children survive.     In the case of education, the relative productivity interpretation is anchored in the large  empirical literature measuring the returns to education at the individual level.  A rough consensus from  this literature is that an additional year of school raises earnings by about 8 percent.  This evidence can  be used to convert differences in learning‐adjusted years of school across countries into differences in  worker productivity.  For example, compared with a benchmark where all children obtain a full 14 years  of school by age 18, a child who obtains only 9 years of education can expect to be 40 percent less  productive as an adult (a gap of 5 years of education, multiplied by 8 percent per year).  Details on the  education component of the HCI are provided in Appendix A2.    In the case of health, the relative productivity interpretation is based on the empirical literature  on health and income, in two steps.  The first step relies on the evidence on health and earnings among  adults.  Many of these studies have used adult height as a proxy for overall adult health, since adult  height reflects the accumulation of shocks to health through childhood and adolescence.  These studies  focus on the relationship between adult height and earnings across individuals within a country.  A  baseline estimate from these studies is that the improvements in overall health that are associated with  an additional centimeter of height raise earnings by 3.4 percent.  However, representative data on adult  height are not widely available across countries.  Constructing an index with broad cross‐country  coverage requires a second step in which the relationship between adult height and more widely‐ available summary health indicators such as stunting rates and adult survival rates is estimated.  Putting  4    the estimates from these two steps together results in a “return” to reduced stunting and a “return” to  improved adult survival rates.  Baseline estimates suggest that an improvement in overall health that is  associated with a reduction in stunting rates of 10 percentage points raises worker productivity by 3.5  percent.  Similarly, an improvement in overall health that is associated with an increase in adult survival  rates of 10 percentage points raises productivity by 6.5 percent.  In countries where data on both  stunting and adult survival rates are available, the average of the improvements in productivity  associated with both health measures is used as the health component of the HCI.  When stunting data  is not available (most commonly for rich countries), only adult survival rates are used.  Details on the  health component of the HCI are provided in Appendix A3    Figure 5 and Figure 6 show the components of the HCI expressed in terms of worker  productivity relative to the benchmark of complete education and full health.  The vertical axis in each  graph runs from zero to one.  The distance between a country’s value and one shows how much  productivity is lost due to the corresponding component of human capital falling short of the benchmark  of complete education and full health.  The benchmark of “complete education” is defined as 14  learning‐adjusted years of school.  The benchmark of “full health” is defined as 100 percent adult  survival and no stunting.   Under the assumptions spelled out in the technical appendix, multiplying  together the three components expressed in terms of relative productivity results in a human capital  index that measures the overall productivity of a worker relative to this benchmark.  The index ranges  from zero to one, and a value of   means that a worker of the next generation will be only  100  percent as she would be under the benchmark of complete education and full health.  Equivalently, the  gap between   and one measures the shortfall in worker productivity due to gaps in education and  health relative to the benchmark.  4.  The Human Capital Index    The overall human capital index is shown in Figure 7.  The units of the HCI have the same  interpretation as the components measured in terms of relative productivity.  Consider for example a  country such as Morocco, which has a HCI equal to around 0.5.  This means that, if current education  and health conditions in Morocco persist, a child born today will only be half as productive as she could  have been relative to the benchmark of complete education and full health.  The HCI exhibits substantial  variation across countries, ranging from 0.3 in the poorest countries to 0.9 in the best performers.  5      All of the components of the HCI are measured with some error, and this uncertainty naturally  has implications for the precision of the overall HCI.  To capture this imprecision, the HCI estimates for  each country are accompanied by upper and lower bounds that reflect the uncertainty in the  measurement of the components of the HCI.  As described in more detail in Section A4.4, these bounds  are constructed by calculating the HCI using lower‐ and upper‐bound estimates of the components of  the HCI.  The resulting uncertainty intervals are shown in Figure 8, as vertical ranges around the value of  the HCI for each country.  These upper and lower bounds are a tool to highlight to users that the  estimated HCI values for all countries are subject to uncertainty, reflecting the corresponding  uncertainty in the components.  In cases where these intervals overlap for two countries, it indicates  that the differences in the HCI estimates for these two countries should not be over‐interpreted since  they are small relative to the uncertainty around the value of the index itself.  This is intended to help to  move the discussion away from small differences in country ranks on the HCI, and towards more useful  discussion around the level of the HCI itself and what it implies for the future productivity of children  born today.    Another feature of the HCI is that it can be disaggregated by gender, for the 126 countries  where gender‐disaggregated data on the components of the index are available.  Gender gaps are most  pronounced for survival to age 5, adult survival, and stunting, where girls on average do better than  boys in nearly all countries.  Expected years of school is higher for girls than for boys in about two‐thirds  of countries, as are test scores.  The gender‐disaggregated overall HCI is shown in Figure 9.  Overall, HCI  scores are higher for girls than for boys in the majority of countries.  The gap between boys and girls  tends to be smaller and even reversed among poorer countries, where gender‐disaggregated data also is  less widely available.    The HCI uses returns to education and health to convert the education and health indicators  into worker productivity differences across countries.   The higher are these returns, the larger are the  resulting worker productivity differences.  The size of the returns also influences the relative  contributions of education and health to the overall index.  For example, if the returns to education are  high while the returns to health are low, then cross‐country differences in education will account for a  larger portion of cross‐country differences in the index.  The information in Figure 5 and Figure 6  provides a sense of the relative contributions of the different components of the HCI.  Learning‐adjusted  years of school range from around 3 to a potential maximum of 14.  This gap implies that children in  countries near the bottom of the distribution of expected years of school will only be 40 percent as  6    productive as future workers as children with complete high‐quality education.  The productivity gaps  associated with differences in health outcomes across countries are somewhat smaller.  Using adult  survival rates as a proxy for overall health, future worker productivity in countries with the worst health  outcomes is about 75 percent of what it could be if children enjoyed full health.  Using stunting rates,  the comparable figure is around 85 percent.    Although different assumptions about the returns to education and health will affect countries’  relative positions in the index, in practice these changes are small since the health and education  indicators are strongly correlated across countries.  This is illustrated in Figure 10, which compares the  baseline index with three alternatives based on different values for the return to health, using adult  survival rates as the health indicator.  The top two panels consider weights based on low‐end and high‐ end estimates from the empirical literature on the returns to height, while the bottom panel arbitrarily  assumes that cross‐country differences in health and education have equally‐sized contributions to  productivity differences (which implies a return to health almost three times as large as the baseline).  In  all cases, the correlation of the baseline index with the index based on alternative weights is very high,  indicating that the precise choice of weights does not matter greatly for countries’ relative positions on  the index.  5.  Connecting the Human Capital Index to Future Income Levels and Growth    The HCI is measured in terms of the productivity of next generation of workers, relative to the  benchmark of complete education and full health.  This gives the units of the index a natural  interpretation:  a value of   for a particular country means that the productivity as a future worker of a  child born today is only a fraction   of what it could be under the benchmark of complete education and  full health.  The relative productivity units of the HCI make it straightforward to connect the index to  scenarios for future aggregate per capita income and growth.  Imagine a “status quo” scenario in which  the expected learning‐adjusted school years and health as measured in the HCI today persist into the  future.  Over time, new entrants to the workforce with “status quo” health and education will replace  current members of the workforce, until eventually the entire workforce of the future has the expected  learning‐adjusted school years and level of health captured in the current human capital index.  This can  be compared with a scenario in which the entire future workforce benefits from complete high‐quality  education and enjoys full health.  Per capita GDP in this scenario will be higher than in the “status quo”  scenario, through two channels:  (a) a direct effect of higher worker productivity on GDP per capita, and  7    (b) an indirect effect reflecting greater investment in physical capital that is induced by having more  productive workers.      Under standard assumptions from the macro development accounting literature (that are  detailed in Appendix A5), projected future per capita GDP will be approximately 1/  times higher in the  “complete education and full health” scenario than in the “status quo” scenario for a country where the  value of the HCI is  .  For example, a country such as Morocco with an HCI value of 0.5 could in the long  run have future GDP per capita in this scenario of complete education and full health that is  approximately 1/0.5 or two times higher than in the status quo scenario.  What this means in terms of  average annual growth rates of course depends on how “long” the long run is.  For example, under the  assumption it takes 50 years for these scenarios to materialize, then a doubling of future per capita  income relative to the status quo corresponds to roughly 1.4 percentage points of additional growth per  year.    6.  Conclusions and Caveats    Like all cross‐country benchmarking exercises, the HCI has limitations.  Components of the HCI  such as stunting and test scores are measured only infrequently in some countries, and not at all in  others.  Data on test scores come from different international testing programs that need to be  converted into common units, and the age of test takers and the subjects covered vary across testing  programs.  Moreover, test scores may not accurately reflect the quality of the whole education system  in a country,  to the extent that tests‐takers are not representative of the population of all students.   Reliable measures of the quality of tertiary education do not yet exist, despite the importance of higher  education for human capital in a rapidly‐changing world.  Data on enrollment rates needed to estimate  expected school years often have many gaps and are reported with significant lags. Socio‐emotional  skills are not explicitly captured.  Child and adult survival rates are imprecisely estimated in countries  where vital registries are incomplete or non‐existent.      One objective of the HCI is to call attention to these data shortcomings, and to galvanize action  to remedy them.  Improving data will take time.  In the interim, and recognizing these limitations, the  HCI should be interpreted with caution.  The HCI provides rough estimates of how current education and  health will shape the productivity of future workers, and not a finely‐graduated measurement of small  differences between countries.  Naturally, since the HCI captures outcomes, it is not a checklist of policy  actions, and right type and scale of interventions to build human capital will be different in different  8    countries.  Although the HCI combines education and health into a single measure, it is too blunt a tool  to inform the cost‐effectiveness of policy interventions in these areas – which should instead be  assessed based on careful cost‐benefit analysis and impact assessments of specific programs.  Since the  HCI uses common estimates of the economic returns to health and education for all countries, it does  not capture cross‐country differences in how well countries are able to productively deploy the human  capital they have.  Finally, the HCI is not a measure of welfare, nor is it a summary of the intrinsic values  of health and education – rather it is simply a measure of the contribution of current health and  education outcomes to the productivity of future workers.      9          Figure 1:  Probability of Survival to Age 5  Probability of Survival to Age 5 1 MNEBLR LVA HRV SVN EST CYP PRT HUN POLCZE ISR LTU ISL FIN JPN ESP KOR ITA SWE AUS DEU AUT NLD BEL FRA GBR DNK TWN SMRNOR IRL CHESGPLUX CUB BIH SRB BHS ATG CHLGRC SVKNZL MLT CAN BHR USA SAU UKR ALB LKA GEO CRIBGR LBN MDV CHN THA RUS MYS ROU URY KAZ ARG TUR OMNHKG ARE MAC QAT KWT BRN ARM BLZ SLV JAM TUN XKXECU PER BRB MEX MKD COL BRAIRNPAN PLW MUSKNA SYC MDA SYRTON NIC HND WSM CPV VCT LCA JOR GRD MNG SLB KGZPSE VNM PRYEGY UZB SUR TUV MAR FJI IDNDZAAZE Probability of Survival to Age 5 VUT GTM PHL TTO KHM FSM BTN GUY DOM IRQ VEN STP BGD NPL TJKMHL BOL DMA NRU RWA IND ZAF BWA MDG SEN KEN COG MMR NAM TLS GAB TKM .95 UGA ZWE GHA MWI KIRTZA YEM PNG SWZ BDI ETH ZMB GMB DJISDN LAO COMAFG LBR MOZ HTI TGO PAK BFA MRT AGO NER GNBGIN CMR LSO COD CIV GNQ SSD BEN .9 NGA MLI SLE CAF TCD .85 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  Probability of survival until age 5 is one minus the under‐5 mortality rate.  Estimates of under‐5 mortality  rates are taken from the UN Inter‐Agency Group on Child Mortality Estimation (www.childmortality.org), and  supplemented with data provided by World Bank staff.  Real GDP per capita adjusted for differences in purchasing  power parity is taken from the Penn World Tables (Feenstra et. al. (2015)), with missing countries filled using data  from World Bank estimates of GDP at PPP.  Graph shows the most recent data for all countries.          10      Figure 2:  Quantity and Quality of Education Data  Panel A:  Expected Years of School By Age 18  Expected Years of School 14 BLRRUS CZE PRT FRA AUT GBR DEU SWE NLD AUS SGP MNG SYCISR SVN JPN LTU FIN NZL ITACAN KOR IRL NOR SRB KAZCYP DNK BELHKG ISL ECU CHN ARG HRV LVA POLMLTBHR OMN USA CHE ARE UKR ALB LKA CUB BRB BGR EST HUN ESP SVK TWN SMRBRN PHL XKX PER CHLGRC VGB KGZ VNM IDNCOL DMA GEO VCT LCA GRD MEX CRI THA MNEMUS TTO ABW SAU KWT MAC LUX QAT BOL GUY UZB PLWTUR ROU MYS 12 TUV MDAWSM BLZ BIHBRA JAM URY NPL KIR GHANIC PRYJOR VEN IRN AZE TKM HTI PSE SLVEGY FJI DZA DOM PANBHS BGD TON ARM MKD ATG TJK KEN LAO BMU VUT MAR SUR LBN NRU PRI MHL TUN STP HND IND 10 ZWE TLS CPV MMR KHM GTM BTN ZAF Years MWICOD BEN SLB CMR SLETGO GMB ZMB PAK COG NAM COMAFG LSO BWA PNGNGA SWZ GAB ETH YEM 8 TZA AGO BDI GNB MDG MOZ SEN CIVSDN GINUGA IRQ RWA BFA SYR CAF MRT 6 MLI DJI NER TCD LBR SSD 4 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Panel B:  Harmonized Test Scores  Harmonized Test Scores 600 SGP JPN KOR HKG TWN RUS KAZEST FIN POL IRL MAC CUB LVASVNCANDNK NLD DEU VNM SRB CZE PRT SWE AUT AUS BEL CHE USA HUN LTUNZL GBR ITA ESP NOR 500 HRVSVKISRFRA BGR CYP ISL LUX UKR Test Scores AZE MUS GRCMLT MYS CHLSYC KEN BIH CHN GAB TUR TTO KHM ROU BHR ARE TJK GEO ARM URY MDA SWZ MNG THA MNE IRN QAT MMR ALB MEX CRI VEN COL ARG OMN BDI KGZ ECU GIN SEN PSE PHL JOR PER NAMBRA SAU 400 BFA SYR HND GTM IDN LBN LKA COM UGA ZWE LSO TUV NIC JAM BWA PAN TZA TGOBEN KIR CMR SDNTON PRY TUN MKD XKX DZA KWT MOZ TLS NPL BGDCIV COG LAO MAR MWI ETH RWASLB PNG IND SLVEGY IRQ MDG AFG VUTZMB DOM HTI GMB MRT PAK GUY ZAF LBR SSD TCD NGA AGO COD YEM SLE MLI GHA 300 NER 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  Expected years of school are calculated using repetition‐adjusted enrollment rates by school level to proxy for  age‐specific enrollment rates up to age 18.  Enrollment rates are taken from the UNESCO Institute for Statistics, and  extensively revised/updated/expanded with estimates provided by World Bank staff.  Harmonized test scores are  taken from Patrinos and Angrist (2018) and are measured in TIMSS‐equivalent units, i.e. a mean of 500 and a standard  deviation of 100 across students in OECD countries.  Real GDP per capita adjusted for differences in purchasing power  parity is taken from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled using data from  World Bank estimates of GDP at PPP. Graph shows the most recent data for all countries.      11      Figure 3:  Expected Learning‐Adjusted Years of School  Expected Learning-Adjusted Years of School SGP JPN KOR 12 RUS FIN HKG CZE SVN DEU CAN NLD AUT SWE IRL KAZPRT EST POL AUS GBR DNK NOR TWN SRB LVA LTUNZL ITA ISRFRA BEL CHE MAC USA CUB HRVCYP HUN ESPISL VNM UKR BGR SYCSVK 10 MLT LUX CHN MUSCHLGRC BHR MNG ARE ALB AZEARG TTO MYS GEO ECU TUR OMN KGZ BIH MEX ROU THA CRI COLMNE QAT MDA PHL LKA PER IRN URY SAU Years IDN VEN 8 KEN TJK PSE ARM XKX JOR BRA KWT TUV NIC JAM PAN KIR NPL KHM PRY MMRGUY LBN DZA MKD ZWE BGD TON HND LAO SLVEGYDOM HTI GTM MAR TUN TLS VUT GAB 6 TGOBEN LSO CMR GHA IND SWZ NAM MWI COM SLB BDI ZMB COG ZAF BWA COD AFGTZA GMB SEN PNG PAK SLE MOZGINUGA ETH SDNNGA MDGBFA YEM CIV SYR AGO IRQ 4 RWA MRT NER MLI TCD LBR SSD 2 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018     Notes:  Learning‐adjusted years of school are measured as expected years of school (top panel of Figure 2) multiplied  by the ratio of each country’s harmonized test score (bottom panel of Figure 2) to a benchmark score of 625,  corresponding to the threshold of advanced attainment set by TIMSS.  Real GDP per capita adjusted for differences in  purchasing power parity is taken from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled  using data from World Bank estimates of GDP at PPP.  Graph shows the most recent data for all countries.      12      Figure 4:  Health Indicators  Panel A:  Adult Survival Rate  Adult Survival Rate 1 CYPITA ISL C HKG HE SGPMAC LBN ISR MLT JPN ESP KOR NZLAUS SWE NLD CAN IRL NOR LUX QAT MAR ALB MDV GRC AUT GBR Adult Survival Rate (Age 15 to 60) SVN BHR DEU DNK BEL CUB TUN CHNIRN CHL PRT CRI CZE FIN FRA SAU ARE KWT BRN XKXBIHDZAMNE MKD BRB TUR HRV ABW PRITWN OMN USA .9 WSM PSE VNM JOR MEX URY SRB ARG PAN SVK CPV ARMECU PER POL EST MYS BGD TON VUT TJK JAM LKA LCA AZE ATG ROU BGR HUN SLB NPL TLS NIC HND UZB DMA PRYEGY COL VEN GRD BRA THAMUS GEO FSM KHM PAK MDA GTM VCTIRQ DOM LVA BHSSYC TTO LTU KGZSYR IND BLZ SLVIDN SUR SEN STP BOL LAO FJIUKR BLR TKM RWA KIR MMRPHL BTN MNGBWA KAZ .8 MDG TZA MRT GUY ETH COMAFGYEM KEN SDN PNG RUS LBR HTI BEN GHA AGO GAB NER BFA GIN DJI COG CODGMB MLI GNB MWI TGO BDI ZMB NAM MHL .7 MOZ UGA SSD ZAF ZWE CMR GNQ TCD NGA CAF SLE CIV .6 SWZ LSO .5 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Panel B:  Fraction of Children Under 5 Not Stunted  Fraction of Children Under 5 Not Stunted 1 LCA CHL AUS KOR USA WSM PRY BRA MKD CRI SRB KWT MDA JAM PSE DOMIRN JPN TON FJI JOR BRB KAZ CHN SUR SYC TUV ARMBIH MNETUR TUN MNG GUY GEODZA THA COL URY TTO KGZ MEX TKM SLV PER VEN OMN BLZ MAR SEN STP NIC BOL LKAGAB Fraction Not Stunted GHA AZE MDV PAN BRN .8 CIVCOG BTN EGY MYS HTI HND NAM ALBIRQ GMB VNM ECUNRU SWZ SLETGO GNB ZWE KEN BFA TJK MRT SYR ZAF GNQ UGA VUT DJI MMR MLI COMSSD BWA LBR GIN SLB CMR KHM BENLSO LAO PHL IDN TZA RWA NPL BGD MWI ETH SDN IND AGO TCD ZMB .6 CAF AFG NER MOZ COD NGA PAK YEM GTM MDG TLS PNG BDI .4 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  Adult survival rates are estimated by the UN Population Division and refer to the fraction of 15 year‐olds who  survive to age 60.  Stunting rates are taken from the WHO‐UNICEF‐World Bank Joint Malnutrition Estimates and refer  to the fraction of children under 5 who are more than two reference standard deviations below the reference median  height for their age.  Data are supplemented with estimates provided by World Bank staff.  The graph reports the  complementary proportion of children who are not stunted.  Real GDP per capita adjusted for differences in  purchasing power parity is taken from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled  using data from World Bank estimates of GDP at PPP.  Graph shows the most recent data for all countries.        13        Figure 5:  Contribution of Education to Productivity  Contribution of Education to Relative Productivity SGP .9 JPN KOR FIN HKG Productivity Relative to Benchmark RUS IRL CZE SVN DEU CAN NLD SWE AUT KAZPRT AUS GBR DNK EST POL LVA TWN FRA NZL NOR .8 SRB LTUITABEL ISR CHE USA MAC CUB HRVCYP HUN ESPISL VNM UKR SVK BGR SYC MLT LUX CHN CHL GRC BHR .7 MNG MUS ARE TTO MYS ALB AZEARG GEO ECU TUR OMN BIH MEX ROU THA CRI COLMNE QAT KGZ MDA PHL LKAPER IRN URY SAU KEN IDN VEN ARM TJK PSE XKX JOR BRA KWT .6 TUV NIC JAM PAN KIR NPL KHM PRY MMRGUY DZA MKD LBN ZWE BGD TON HND LAOSLVEGYDOM HTI GTM MAR TUN TLS VUT GHA IND SWZ NAM GAB TGOBEN MWI COM CMR LSO .5 BDI SLB ZMB COG ZAF BWA COD AFGTZA GMB SEN PNG PAK SLE MOZ GINUGA ETH SDN MDG BFA YEM CIV SYRNGA AGO IRQ RWA MRT NER MLI TCD .4 LBR SSD 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  This graph reports the contribution of cross‐country differences in learning‐adjusted years of school to cross‐ country differences in worker productivity.  The vertical axis measures the productivity of a worker relative to the  benchmark of complete education.  Differences in years of school are converted to productivity differences using  estimates of the returns to school detailed in Appendix A2.  Real GDP per capita adjusted for differences in purchasing  power parity is taken from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled using data  from World Bank estimates of GDP at PPP.  Graph shows the most recent data for all countries.      14      Figure 6:  Contribution of Health to Productivity  Panel A:  Based on Adult Survival Rates  Contribution of Health to Relative Productivity (ASR) 1 CYPITA ISL CHE HKG SGPMAC ISR MLT JPN ESP SWE AUS NLD IRL NOR Productivity Relative to Benchmark MAR ALB LBN MDV GRC KOR NZLCAN GBR AUT BHR DEU DNK LUX QAT SVN CUB TUN CHNIRN CHL PRT CRI CZE BEL FIN FRA SAU ARE KWT BRN BIHDZA XKX MKD MNE BRB TURHRV ABW PRITWN OMN JOR MEX URY SVK USA PSE WSM CPV ARM VNM PER ECU SRB ARG PAN POL EST MYS BGD TON VUT TJK JAM LKA LCA AZE ATG ROU BGR HUN SLB NPL TLS NIC HND UZB DMA PRYEGY GEO COL VEN GRD BRA THAMUS PAK GTM VCT DOM IRQ LVA BHSSYC .9 FSM KHM MDA BLZIDN SURBLR TTO LTU KGZ SEN STP SYR IND BOL SLV RWA KIR LAO MMR FJIUKR TKM MDG TZA MRT PHL BTN MNGBWA KAZ GUY ETH COMAFGYEM KENPNG SDN RUS LBR HTI BEN GHA AGO GAB NER COD BFA GIN DJI COG GMB MLI MWI GNB TGO BDI ZMB MHL NAM MOZ SSDUGA ZAF ZWE CMR GNQ .8 TCD NGA CAF SLE CIV SWZ LSO .7 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Panel B:  Based on Stunting Rates  Contribution of Health to Relative Productivity (Stunting) 1 LCA CHL AUS KOR USA WSM PRY BRA MKD CRI SRB KWT MDA JAM DOMIRN JPN Productivity Relative to Benchmark PSE JOR BRB TON FJI BIH CHN SYC KAZ ARM SUR MNETUR TUV TUN MNGTHA COL URY TTO GUY GEO DZAMEX TKM KGZ SLV PERVEN .95 BLZ MAR OMN SEN STP NIC BOL LKAGAB AZE GHA MDV PAN BRN COG CIV BTN EGY MYS HTI HND IRQ NAM ALB GMB VNM ECUNRU SWZ SLETGO GNB ZWE KEN BFA TJK MRT SYR ZAF GNQ UGA VUT DJI MMR .9 MLI COM SSD SLB CMR BWA LBR GIN LSO KHM LAO PHL IDN BENTZA RWA NPL BGD MWI ETH SDN IND AGO CAF TCD AFG ZMB NER MOZ COD NGA PAK .85 YEM GTM MDG TLS PNG BDI .8 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  This graph reports the contribution of cross‐country differences in health outcomes to cross‐country  differences in worker productivity.  The vertical axis measures the productivity of a worker relative to the benchmark  of full health.  Differences in health outcomes are converted to productivity differences using estimates of the returns  to health detailed in Appendix A3.  Real GDP per capita adjusted for differences in purchasing power parity is taken  from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled using data from World Bank  estimates of GDP at PPP.  Graph shows the most recent data for all countries.    15      Figure 7:  The Human Capital Index  Human Capital Index 1 Productivity Relative to Benchmark Singapore Korea, Rep. .8 Germany United States Russian Federation Vietnam China Turkey Thailand .6 Colombia Philippines Indonesia Kenya Morocco India Malawi Benin South Africa .4 Mozambique Côte d'Ivoire Chad .2 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018   Notes:  This figure reports the Human Capital Index.  The vertical axis measures productivity relative to the  benchmark of complete education and full health.  A value of  on the vertical axis means that the productivity as a  future worker of a child born today is only  100 percent what it would be in the benchmark of complete education  and full health.  Real GDP per capita adjusted for differences in purchasing power parity is taken from the Penn World  Tables (Feenstra et. al. (2015)) , with missing countries filled using data from World Bank estimates of GDP at PPP.   Graph shows the most recent data for 157 countries.  Selected countries are labelled for illustrative purposes.        16      Figure 8:  The Human Capital Index, With Uncertainty Intervals  Human Capital Index 1 (With Uncertainty Intervals) Productivity Relative to Benchmark Singapore Korea, Rep. .8 Germany United States Russian Federation Vietnam China Turkey Thailand .6 Colombia Philippines Indonesia Kenya Morocco India Malawi Benin South Africa .4 Mozambique Côte d'Ivoire Chad .2 6 8 10 12 Log Real GDP Per Capita Note: Dataset version 21 Sept 2018 Note: vertical range indicates uncertainty interval around HCI estimate     Notes:  This figure reports the Human Capital Index.  The vertical axis measures productivity relative to the  benchmark of complete education and full health.  A value of  on the vertical axis means that the productivity as a  future worker of a child born today is only  100 percent what it would be in the benchmark of complete education  and full health.  Uncertainty intervals around estimates are shown as vertical ranges for each country.   Real GDP per  capita adjusted for differences in purchasing power parity is taken from the Penn World Tables (Feenstra et. al.  (2015)) , with missing countries filled using data from World Bank estimates of GDP at PPP.  Graph shows the most  recent data for 157 countries.  Selected countries are labelled for illustrative purposes.        17      Figure 9:  Gender Differences in the Human Capital Index  Human Capital Index (Gender Disaggregated) 1 Productivity Relative to Benchmark Singapore Korea, Rep. .8 Germany United States Russian Federation China Turkey .6 Thailand Colombia Indonesia Philippines Morocco India Benin .4 South Africa Côte d'Ivoire Chad .2 6 8 10 12 Log Real GDP Per Capita at PPP Note: Dataset version 21 Sept 2018 Legend: Square -- Male, Line -- Female   Notes:  This figure reports the Human Capital Index.  The vertical axis measures productivity relative to the  benchmark of complete education and full health.  A value of  on the vertical axis means that the productivity as a  future worker of a child born today is only  100 percent what it would be in the benchmark of complete education  and full health.  Real GDP per capita adjusted for differences in purchasing power parity is taken from the Penn World  Tables (Feenstra et. al. (2015)) , with missing countries filled using data from World Bank estimates of GDP at PPP.   Graph shows the most recent data for 131 countries where gender‐disaggregated data for all of the HCI components  is available.  Selected countries are labelled for illustrative purposes.      18        Figure 10:  Effect of Changing Weight on Health in the Human Capital Index  1 Human Capital Index: Robustness to Weights 1 High Weight on Health Low Weight on Health .8 .8 .6 .6 .4 .4 Corr=.998 Corr=.997 .2 .2 .2 .4 .6 .8 1 .2 .4 .6 .8 1 Baseline Weight on Health Baseline Weight on Health Equal Weight on Health and Education .4 .6 .8 1 Corr=.993 .2 .2 .4 .6 .8 1 Baseline Weight on Health Note: Dataset version 21 Sept 2018   Notes:  This graph shows the effect of changing the weights on the health and education components of the HCI.  In  each panel the horizontal axis corresponds to the HCI with baseline weights.  In the top‐left (top‐right) panel the  vertical axis corresponds to the HCI assuming a low‐end (high‐end) estimate for the return to health from the  empirical literature, as discussed in Appendix A3.  The bottom‐left panel assumes a much larger value for the return  to health that generates the same gap in productivity between best and worst performers as is observed between the  best and worst performers in learning‐adjusted years of school.  Real GDP per capita adjusted for differences in  purchasing power parity is taken from the Penn World Tables (Feenstra et. al. (2015)) , with missing countries filled  using data from World Bank estimates of GDP at PPP.  Graph shows the most recent data for all countries.       19    Technical Appendix:  Detailed Methodology for The Human Capital Index    A1:  Basic Framework    A2:  Education    A2.1  Data on Expected Years of School    A2.2 Harmonizing Test Scores    A2.3  Adjusting Expected Years of School for Quality    A2.4  Returns to Education    A3:  Health    A3.1  Basic Methodology    A3.2  Estimates of the Return to Height    A3.3  The Relationship Between Adult Height and Adult Survival    A3.4  The Relationship Between Adult Height and Stunting      A4:  The Human Capital Index    A4.1  Putting the Pieces Together    A4.2  Robustness to Alternative Weights    A4.3  Disaggregation by Gender    A4.4  Uncertainty Intervals for the HCI and Its Components    A5:  Linking the Human Capital Index to Future Growth Scenarios      20    A1.  Basic Framework      This section sets out a simple framework used by the development accounting literature to  measure human capital and uses it to motivate the Human Capital Index (HCI).2  This literature begins  from the observation that the productivity of an individual worker is higher the more educated she is  and the healthier she is.  This gain in productivity represents the contributions of health and education  to her human capital.     Let   represent the years of school of an individual worker  , and let   be a measure of her  health.  The human capital of a worker is:    (1)  ℎ       Section A2 discusses how years of school   are measured and adjusted for differences in quality as  reflected in performance on international student achievement tests.  Section A3 discusses the mapping  from unobserved “latent” health   to observable health indicators. The parameters  and  represent  the “returns” to an additional unit of education and health.  For example, when education is measured  as years of school, this formulation implies that an additional year of school raises the human capital of  the worker by 100  percent.   As detailed in Sections A2 and A3, plausible values for  and  can be  drawn from the large microeconometric literature that has estimated returns to education and health  using individual‐level data.      The expected future human capital of a child born today is:                                                                 2  Klenow and Rodriguez‐Clare (1997) and Hall and Jones (1999) are early examples of the development accounting  approach, and Caselli (2005) and Hsieh and Klenow (2012) provide surveys. See also Caselli (2014) for an  application of this methodology to Latin America, commissioned by the LAC region of the World Bank. The  discussion of the contribution of health to human capital draws heavily on Weil (2007) and Ashraf, Lester and Weil  (2009).  Galasso and Wagstaff (2016) use the development accounting approach to assess the macroeconomic  costs of stunting.    21    (2)  ℎ     where   and   represent her expected future education and health;   is the probability that a child  born today survives; and   represents the Next Generation of workers. 3  Multiplying by  captures the  loss in future productivity per child born today due to premature mortality, since children who do not  survive do not grow up to become productive adults.  The survival probability  is the complement of  the under‐5 mortality rate.4  As discussed below, expected future education and health are measured  based on the current outcomes.  For example, expected future education will be measured as the  number of years of school a child progressing through the education system is likely to obtain given  prevailing enrollment rates at different levels.  Similarly, expected future health will be measured under  the assumption that current health conditions prevail into the future.     Human capital in Equation (1) expresses human capital in units of productivity relative to a  worker with  0, in which case ℎ 1.   To express the HCI in more intuitive units, rescale  Equation (1) by dividing by a benchmark level of human capital corresponding to complete education  and full health.  Let ∗ ,  ∗  and  ∗  represent these benchmark values.  For survival, a natural benchmark  is ∗ 1.  For years of school, the benchmark is  ∗ 14 years of school, corresponding the maximum  possible number of years of school achieved by age 18 by a child who starts school at age 4.  For health  the natural benchmark corresponding to full health is  ∗ 1.     With this notation, the HCI is:                                                                 3  Formally, let ℎ  represent human capital at some future date  . Expected future human capital is given by  ℎ , where  is the probability a child does not survive to  become a future worker, in which case her human capital as a future worker does not materialize.  The first  equality requires the assumption of independence between education and health outcomes across individuals, and   and   should be interpreted as expectations conditional on survival (and assuming that human  capital conditional on not surviving is zero).  The second inequality is due to the convexity of the human capital  function.  Since only the “likely future values” of  health and education,    and  , are observable, and  not the entire distribution of possible future outcomes which would be required to calculate   and  , the last term serves as a lower bound on expected future human capital.  Naturally, given the convexity  of the human capital function, a higher variance of education and health across individuals, and a higher  covariance between the two, increases the gap between the lower bound and the expectation.   To keep notation  simple,   and   denote the likely future values   and   that represent the expected education  and health of the next generation of workers.    4  Data on under‐5 mortality are produced by the UN Child Mortality Estimates.  Most of the cross‐country variation  in mortality under 5 is due to cross‐country variation in under‐1 mortality rates.  22    ∗ ∗ (3)    ∗     The HCI is the product of three easily‐interpretable components, each measuring productivity  relative to the benchmark of full health and complete education.  The first term,  ∗, captures forgone  future productivity due to child mortality, since children who do not survive never become productive  adults.  As a result, the average productivity as a future worker of a child born today is reduced by a  factor equal to the survival rate, relative to the benchmark where all children survive.  The second term,  ∗ , reflects foregone future productivity due to children completing less than a full 14 years of  ∗ school.  The third term,  , reflects the reduction in future worker productivity due to poor  health.  Multiplying these three terms together gives the overall productivity of a worker relative to the  benchmark of complete education and full health.    This approach is closely linked to standard measures of the average human capital per worker of  the current workforce that have been widely used in the development accounting literature:    (4)  ℎ     where ℎ  represents the average human capital of the current workforce, and   and   represent  the average levels of education and health in the current workforce.  The only difference between this  measure and the expected human capital of the next generation in Equation (2) is that the term  reflecting the probability of survival is not required.  This is because the measure of human capital of the  current workforce measures the average human capital of workers who are currently living.      While measures of the human capital stock like those in Equation (4) are standard in the  development accounting literature (see for example Weil (2008)), they are less well suited to the  communications and advocacy purpose of the HCI.  This is because measures of the human capital of the  existing workforce – and most particularly the education component, reflect the educational  opportunities that were available to current workers in the past when they were school‐aged children,  and so now are largely beyond the influence of current and future policy interventions.  Instead, the HCI  measures how current health and education outcomes – that are amenable to improvement through  current and future policy efforts – shape the likely future human capital of children born today.    23      Measures of the monetary value of human capital based on the present value of future earnings  of individuals, analogous to estimates of the value of physical capital as the present value of future  returns, also exist.  Naturally, these measures are conceptually closely related.  Suppose for example  that log wages of individual   at some future time  are given by a health‐augmented Mincer equation  like   , where   represents future trend growth in wages for the individual.   Treating the unskilled wage as the numeraire, human capital measured as the present value of future  wages is simply  , where   represents the discount rate, and ℎ  is the measure of individual human  capital in Equation (1).  Human capital measures along these lines have a long history (see for example  Jorgenson and Fraumeni (1998)), and are extensively discussed in the context of satellite national  accounts in UN (2016).  Measures of human capital along these lines in a cross‐country setting have  been developed since 2012 in the United Nations University “Inclusive Wealth Index” study (UNU  (2012)), as well as in the latest edition of the World Bank’s “Changing Wealth of Nations” report (World  Bank (2018)).  The key incremental difficulty in constructing these measures relative to measures of ℎ  is  coming up with plausible measures of future earnings growth,  . Because the difference between the  growth rate and the discount rate is small and enters in the denominator of this measure, small changes  in assumed growth rates are magnified into large changes in measured human capital.5                                                                     5  For example, if the discount rate is five percent, changing the assumed growth rate from three to four percent  per year has the effect of doubling the measured human capital stock.  24    A2:  Education  A2.1  Data on Expected Years of School    Expected number of years of school () that a child who starts school at age 4 would attain by  her 18th birthday is calculated using the methodology described in UNESCO et. al (2014), Section 2.2 and  Annex 2.2.  Conceptually, expected years of school achieved by age  is simply the sum of age‐specific  enrollment rates over all ages in this age range, i.e.     (5)      where   is the enrollment rate of children aged . Unfortunately however, age‐specific enrollment  rates are not systematically available for a broad cross section of countries.  Instead, more readily‐ available data on enrollment rates by level of school are used to approximate enrollment rates in  different age brackets.  Specifically, pre‐primary enrollment rates approximate the age‐specific  enrollment rates for 4 and 5 year‐olds; the primary rate approximates for 6‐11 year‐olds; the lower‐ secondary rate approximates for 12‐14 year‐olds; and the upper‐secondary rate approximates for 15‐17  year‐olds.  Naturally, cross‐country definitions in school starting ages and duration of different levels of  school imply that these will only be approximations to the number of years of school a child can expect  to complete by age 18.      The ideal measure of enrollment rates for this calculation is the “total net enrollment rate”  (TNER), which measures the fraction of children in the theoretical age range for a given level of school,  who are in school at any level.  For example, if the theoretical age range for lower secondary school is 12  to 14 years, then the TNER measures all children aged 12 to 14 who are enrolled in any level of school as  a fraction of all children aged 12 to 14.  In this way, the TNER best approximates the age‐specific  enrollment rates for ages 12 through 14 since it captures the enrollment status of all 12 to 14 year‐olds,  irrespective of what level of school they are in.  Unfortunately however, data on TNER are missing for  many countries and years in the UNESCO database, and, depending on the country and year, one or  more of three other enrollment rates are more widely available.  These are (i) the “adjusted net  enrollment rate” (ANER), measuring the fraction of children in the theoretical age range for a given level  25    of school who are in school at that level or the level above; (ii) the “net enrollment rate” (NER),  measuring the fraction of children in the theoretical age range for a given level of school who are in  school at that level; and (iii) the “gross enrollment rate” (GER), measuring the number of children of any  age who are enrolled in a given level, as a fraction of the number of children in that age range.       To maximize country coverage, the following procedure populates the enrollment rate series  used to calculate expected years of school:   Available data for all four enrollment rates for all four levels of school were retrieved, combining  information from the three external data platforms maintained by the UNESCO Institute of  Statistics.  For pre‐primary, TNERs are not reported as there is no level below pre‐primary, and  ANER is available only for the age corresponding to one year before the official start of primary  school.     Gaps in each enrollment rate for each level and country were filled by taking the most recently  available data, going backwards up to 10 years for each country‐year observation.   Within each level, the preferred enrollment rate available in the filled‐in data as of 2017 was  chosen, in the following order of priority:  TNER, ANER, NER, and GER.  The filled‐in series for  this enrollment rate was then used for all years for this level of school.  Note that in some cases  differences in availability of enrollment measures means that different enrollment rates are  used for different levels of school.  However, with a level for a given country, same type of  enrollment rate is used.   Data on repetition rates for primary, lower‐secondary, and upper‐secondary school were  retrieved, and filled in with up to 10 years of lags in the same way as for enrollment rates.   These are needed to adjust enrollment rates obtained in the previous step for repetition.   Adjusting for repetition is important because failing to do so would count students repeating a  grade as gaining an additional year of school, and in some school systems at some levels  repetition rates are as high as 25 percent.  Data on repeaters is available for most countries  included in the HCI;  However for a few countries where data on repetition is not reported by  UNESCO, out of necessity repetition rates are assumed to be zero.   Finally, in a few cases where only GERs were available and repetition‐adjusted GERs exceeded  100 percent.  These are topcoded at 100 percent.           26      The resulting measure of expected school years approximates the number of years of school  that a child can expect to attain by her 18th birthday if she starts school at age 4, for a maximum of 14  years.  Data on this measure of expected years of school are reported in the top panel of Figure 2.    Conceptually this calculation corresponds to the measure of “school life expectancy” (SLE)  calculated by UNESCO Institute for Statistics.6  However, the implementation here differs in that UIS  uses gross enrollment rates to calculate school life expectancy, whereas here total net enrollment rates  are used wherever possible.  This reflects a tradeoff.  On the one hand, gross enrollment rates are more  widely available and typically have longer time‐series coverage in the UIS data.  On the other hand, total  net enrollment rates conceptually correspond more closely to the age‐specific enrollment rates in  Equation (5).  A further reason to use total net enrollment rates is that for some countries the  repetition‐adjusted gross enrollment rates reported by UNESCO are – sometimes implausibly – well  above 100 percent.  While gross enrollment rates (the number of students enrolled at a given grade  level as a fraction of children of the theoretical age for that grade level) can in principle exceed 100  percent if some children start school early or late, it is difficult for this timing effect to generate gross  enrollment rates that are persistently much above 100 percent.7  A practical consequence of using gross  enrollment rates above 100 percent to calculate SLE is that SLE can then exceed the statutory duration  of school.   This is the case for primary and secondary SLE as reported by UIS for about one‐quarter of                                                               6  When more granular data such as enrollment rates by age are available, UNESCO uses them to calculate SLE.   When they are not, UNESCO calculates SLE using school level‐specific gross enrollment rates (which are the most  widely available in the UNESCO data), using Equation (5).  Whether age‐specific or level‐specific enrollment rates  are used makes little practical difference for the calculations.  Using the data on gross enrollment rates and  duration by level of school as reported by UNESCO in Equation (5), and restricting attention to primary and  secondary, reproduces the UNESCO estimates of SLE for primary plus secondary almost exactly.    7  To see why, consider an education system in which a fraction   of each cohort of students start school at the  official age; a fraction   starts one year late and therefore are above age‐for‐grade; and a fraction 1   students do not start school at all.   Gross enrollment in first grade at time  is the number of students in first grade  divided by the cohort of children aged 6 who are supposed to be in first grade, i.e.    where   is the number of children in the cohort of new six year‐olds at time .  With a stationary population    and stationary starting age shares   and  , it follows that  1, even though there could be  many over‐aged children in the grade.  With stationary starting age shares the only way to generate  1 is if   is close to one (i.e. near‐universal enrollment) and   i.e. a shrinking school‐aged population.  The  other way to generate  1 is with time‐varying starting age shares.  If for example there is a big expansion in  enrollment to include children who should have started in the previous year but did not, then with stable cohort  size it follows that   and it is possible that  1  if enough children from the previous  cohort who did not enroll at the correct age now enroll, i.e. if    is large enough.   One factor contributing to  high gross enrollment rates in some countries is that they can include adults who are returning to complete  secondary school, which can inflate gross enrollment rates beyond 100 percent, making it a poor proxy for  enrollment rates among 14‐17 year‐olds  27    countries in 2015.  In contrast, here enrollment rates below 100 percent are used at all levels, with the  result that the maximum possible expected years of school is fixed at 14 years.  A2.2  Harmonizing Test Scores    The school quality adjustment is based on a new large‐scale effort to harmonize international  student achievement tests from several multi‐country testing programs.  A detailed description of the  test score harmonization exercise is provided in Patrinos and Angrist (2018).  This paper updates and  expands the dataset described in Altinok, Angrist, and Patrinos (2018).  This earlier dataset harmonized  scores from three major international testing programs (Trends in International Maths and Science  Study (TIMSS), Progress in International Reading Literacy Study (PIRLS), and Programme for International  Student Assessment (PISA)), as well as three major regional testing programs (Southern and Eastern  Africa Consortium for Monitoring Educational Quality (SACMEQ), Program of Analysis of Education  Systems (PASEC), and Latin American Laboratory for Assessment of the Quality of Education (LLECE)).   Patrinos and Angrist (2018) subsequently update this dataset with more recent rounds of PIRLS, PASEC  and SACMEQ, and also substantially expand the cross‐national coverage of the database by including  Early Grade Reading Assessments (EGRA).  The expanded dataset covers over 160 countries.      Test scores from these different testing programs can be converted into common units (or  “harmonized”) using a variety of methodologies, as described in detail in Patrinos and Angrist (2018) and  Altinok, Angrist and Patrinos (2018).  In the version of the data used here, the numeraire units are those  of the international standardized achievement tests. This includes TIMSS (for math and science) and  PISA and PIRLS (for reading, at secondary and primary level, respectively). These tests are expressed in  units with a mean of 500 and a standard deviation across students of 100 points.  The harmonization  method is based on the ratio of country‐level average scores on each program to the corresponding  country‐level scores in the numeraire testing program, for the set of countries participating in both the  numeraire and the other testing program.  For example, consider the set of countries that participate in  both PISA and TIMSS assessments.  The ratio of average PISA scores to average TIMSS scores for this set  of countries provides a conversion factor for PISA into TIMSS scores, that can then be used to convert all  countries’ PISA scores into TIMSS scores.  Altinok, Angrist and Patrinos (2018) and Patrinos and Angrist  (2018) refer to the set of common countries as the “doubloon countries”, the resulting conversion factor  as the “doubloon index”, and the test scores in common units as “harmonized learning outcomes  (HLOs)”.  In the version of the data used here, the doubloon index is calculated pooling all doubloon  observations between 2000 and 2017 and therefore is constant over time.  This ensures that within‐ 28    country over‐time fluctuations in harmonized test scores for a given testing program reflect only  changes in the test scores themselves, and not changes in the conversion factor between tests.8    Harmonization is done at the subject x grade level.  The country‐level test scores used in the HCI average  across subjects and grades.    In most cases, the tests are designed to be nationally representative.  There are however some  notable cases where they are not, including:     In a number of countries, EGRA assessments are not nationally‐representative and are identified  as EGRANR in the data documentation.     In India, the 2009 PISA was administered in two states (Himachal Pradesh and Tamil Nadu).   However, a comparison with state‐level scores for all of India in the 2012/2013 National  Achievement Survey (NAS) suggests that the average NAS score for these two states is quite  similar to the national average NAS score, indicating that the 2009 PISA scores probably are  roughly representative of India as a whole.9   PISA scores for China in 2009 and 2012 are based only on reported data for Shanghai, and in  2015 for Beijing, Shanghai, Jiangsu and Guangdong (B‐S‐J‐G).  Shanghai and B‐S‐J‐G are both  considerably richer than China as a whole.  Since test scores tend to improve with income within  and across countries, reported PISA scores are unlikely to be nationally representative.   Corroborating evidence can be found in Gao et. al. (2017) who implement PIRLS assessments in  Shaanxi and rural Jiangxi and Guizhou provinces, the latter being among the poorest areas in  China.  As noted in Gao et. al. (2017), test scores in these areas are among the lowest PIRLS  scores observed globally.  Extrapolations using average household per capita income using (a)  Shanghai and B‐S‐J‐G PISA scores, and (b) PIRLS scores in Gao et. al. (2017), result in similar  estimates of nationally‐representative test scores.  Converting these extrapolations to HLO units  and averaging them together results in an HLO of 456 (as opposed to 532 in B‐S‐J‐G alone).  This  extrapolated estimate is used as the HLO for China.10                                                                 8  The one exception to this is the 2007 and 2014 PASEC rounds, which were not designed to be intertemporally  comparable, and for which different doubloon countries are available in 2007 and 2014.  9  See the appendix of Patrinos and Angrist (2018) for details.  10  See the appendix of Patrinos and Angrist (2018) for details.  29      The next step is to create a single time series of harmonized learning outcomes (HLOs) for each  country, combining HLOs from different testing programs in a way that balances coverage and  comparability over time.  This is done in three main steps:   International Testing Programs:  Data from TIMSS and PIRLS are combined into a single testing  program, recognizing that both tests are carried out by the same parent organization, the  International Association for the Evaluation of Educational Achievement (IEA), and use common  units.  TIMSS assessments have been carried out in 2003, 2007, 2011 and 2015, and PIRLS in  2001, 2006, 2011, and 2016.  PIRLS assessments are shifted to the nearest TIMSS year, and then  a combined TIMSS/PIRLS score is created by averaging the two (for countries participating in  both), and otherwise using whichever of the two is available.  TIMSS/PIRLS and PISA scores are  chosen by taking whichever of the two is available for each country‐year, and the average of the  two in 2003 and 2015 (when the two testing programs were both carried out in the same year,  for the set of common countries).  The end result is that the HLO data as used in the HCI has 175  country‐year observations between 2000 and 2017 based on TIMSS/PIRLS, 278 country‐year  observations based on PISA, and 68 country‐year observations based on an average of PISA and  TIMSS/PIRLS.   Regional Testing Programs (LLECE, SACMEQ, PASEC).  Scores from these programs are used in  the combined HLO measure only for those country‐year observations where no data from  international testing programs is available.  This adds 25 country‐year observations from LLECE,  37 from SACMEQ, and 30 from PASEC.  Six country‐year observations from LLECE and one  observation from PASEC are dropped because the occur in country‐years for which an  international test is available.   Early Grade Reading Assessment (EGRA).  The HLO database contains data from 75 EGRA  assessments carried out in 48 low‐ and middle‐income countries.   There are 27 countries where  EGRA is the only available international test, and so for these countries the HLO simply is based  on the EGRA assessment.  The remaining 22 countries also have data from one of the major  international and/or regional testing programs.  For these countries, EGRA data are included in  the time series of test scores for the country only if (a) the additional test substantially  expanded the time series coverage of the combined HLO, and (b) the additional test did not  30    create an implausibly large change in the combined HLO score.  The end result of this step is to  add 54 country‐year observations based on EGRA assessments.11    Finally, the HLO dataset includes one observation for Sri Lanka in 2013 based on the national  assessment, which was linked to TIMSS.  These steps result in a combined HLO score with 668 country‐ year observations covering 162 countries.  The median country has three time series observations  between 2000 and 2017.  The HCI is based on a cross‐section of test scores in 2017, taking the most  recent test score available in the period 2006‐2017.  For 133 of these countries, the most recently‐ available test score falls in the five‐year period 2013‐2017.  A cross‐section of most recently‐available  test scores are displayed in the bottom panel of Figure 2.  Test scores range from just below 600 in the  best‐performing countries to just above 300 in the worst‐performing countries.  A2.3  Adjusting Expected Years of School for Quality    The next step is to transform test scores into a quality‐adjustment factor for years of school.   There are (at least) five options for doing so:12     The first is to exploit the fact that in some settings the same test is administered to children in  different grades.  If children in different grades are on average similar in their abilities, the  difference in average performance between children in the two grades can be interpreted a  measure of the “productivity” of the additional year in terms of higher test scores.  This gradient  can then be compared across countries to infer cross‐country differences in average school  quality.  One application of this is done directly by PISA, and exploits the fact that PISA tests are  administered to children at age 15, irrespective of what grade they are in.  While most PISA test  takers are in 9th or 10th grade (depending on the country), some are in earlier or later grades.   Comparing the difference in PISA scores across grades within a country provides an estimate of  how many PISA points are gained as a result of one more year of school in that country.                                                                11  There is one exception to the rule of giving preference to international assessments over EGRAs.  This is for  Yemen, which participated in TIMSS in 2007 and 2011.  Test scores in both years were extremely low (low 200s),  and TIMSS caveated the results, noting that too few students achieved minimum performance standards for the  country‐level average scores to be reliably estimated.  These TIMSS observations are droped and the 2011 EGRA  for Yemen is used instead.  12  For an early effort to incorporate cross‐country differences in test scores into measures of human capital, see  Caselli (2005), Section 4.2.  Studies such as Hanushek and Woessman (2010) have incorporated tests scores into  cross‐country growth regressions.  Hanushek, Ruhose, and Woessman (2015) use test scores as a measure of  cognitive ability in a development accounting exercise across US states, and Caselli (2014) performs a similar  exercise across countries in Latin America.    31    Estimates from the latest PISA round suggest that on average roughly 30 PISA points are in this  sense equivalent to an additional year of school.13  This would imply that the observed 300 point  range in test scores in the data correspond to  10 years of school.  A drawback of this  approach is that it does not control for selection.  For example, if high‐ability students are more  likely to be in 10th grade than 9th grade at age 15, then this calculation overstates the return in  terms of PISA points to an additional year of school.  A conceptually‐related approach is taken  by Kaarsens (2014), who exploits the fact that the 1995 TIMSS round was administered to some  3rd and 7th grade students as well as the usual 4th and 8th grade students for this program.  He  finds that a year of school in the US is worth approximately three times as much as in low‐ income countries.     Evans and Yuan (2017) suggest a related approach which instead looks at slope of the  relationship between test score obtained by adults and the number of years of education they  obtained as children.  The main advantage of this approach is looks at the effect of education on  learning over a wider range than just the one‐year comparisons in the PISA approach, and also  controls for selection issues.  They find that a one standard deviation improvement in Skills  Towards Employability and Productivity (STEP) scores – a testing program for adults in  developing countries – is equivalent to between 5 and 7 years of school. The difficulty in  applying these results in this context is that it is unclear how to map standard deviations of STEP  scores into standard deviations of PISA scores.  Under the somewhat arbitrary assumption that  the standard deviations are the same, their results would imply that the 300 PISA point  differences (or 3 standard deviations) observed across countries translate into a very large gap  of 15 to 21 years of school.14   A third approach is suggested in World Bank (2017) and developed further in Filmer, Rogers,  Angrist, and Sabarwal (2018). They propose scaling actual average school years in a country by  the ratio of the country’s test score to a benchmark value of top performance, and explain how  that this scaling factor can be justified under the assumption that grade‐learning trajectories are  linear through the origin.  A natural benchmark value for top performance is a TIMSS score of  625, which corresponds to the TIMSS “advanced” international benchmark.  Since country‐level  average test scores range from around 300 in the worst‐performing countries to 600 in the best‐                                                              13  See OECD (2016), p. 64.   14  A similar exercise is done in Hanushek and Zhang (2009), who use adult test scores for 14 mostly OECD countries  and relate them to years of educational attainment of the test takers.  32    performing countries, the ratio of test scores relative to the leader is 300/625=0.48 in the worst‐ performing countries.  This implies that a year of school in the worst‐performing countries is  “worth” only about half as much as a year of school in the best‐performing countries.   A fourth approach is suggested in Schoellman (2012), who relates estimated returns to  education of immigrants in the United States to test scores in their countries of origin, and finds  a strong positive relationship.  He then introduces this relationship into a model in which agents  take education quality into account when choosing their investment in years of school.  His  estimates suggest that the return to a year of school needs to be doubled relative to standard  benchmarks to take into account the fact that in countries where school attainment is low, this  reveals an endogenous response to low quality.  While this argument does not provide a direct  mapping from individual country test scores, it suggests a factor‐of‐two quality adjustment  which is similar to World Bank (2017) and Filmer, Rogers, Angrist, and Sabarwal (2018).     A fifth approach can be anchored in the small literature that has jointly estimated returns to  years of school and returns to cognitive ability, where cognitive ability is proxied using test  scores.  For example, Hanushek and Yang (2009) estimate Mincerian regressions of log earnings  on education and test scores and find a return per year of school of 10 percent, and a return per  standard deviation of test scores of 10 to 20 percent (as cited in Hanushek, Ruhose and  Woessmann (2015)) .  The ratio of these two estimated returns implies that one standard  deviation of test scores is equivalent to 1 to 2 years of school in terms of its effects on earnings.   Under the somewhat arbitrary assumption that a standard deviation of test scores in the  Hanushek and Yang (2009) study is equivalent to a PISA standard deviation of 100, this would  imply that the 300 PISA point gap in test scores between the worst and best performing  countries is equivalent to approximately 3 to 6 years of school.  At the other extreme, Caselli  (2014) cites estimates in Vogl (2014) that suggest a return to school of 7.2 percent and a return  per 100 PISA points of 1.4 percent, suggesting that the 300 PISA point gap in test scores across  . countries is equivalent to only about  3 0.6 years of school.  .   Absent a clear empirical consensus on the equivalence between test scores and years of school,  the HCI calculations in this document use the mapping from test scores to quality differences suggested  by Filmer, Rogers, Angrist, and Sabarwal (2018) because of its simplicity and ease of communication.   The calculations are implemented using the most recently available test scores for each country.  This  33    amounts to the assumption that the quality of education that children will receive in the future is the  same as what is reflected in the most recently‐observed test scores.    It is also worth noting that for most countries, test scores are observed only at one grade level.   This means that out of necessity quality measured at one grade level is assumed to be representative of  the entire school system.  This assumption however may however be questionable in countries where  there is a lot of attrition as students move to higher grades.  In this case, test scores observed in later  grades reflect an element of selection, to the extent that lower‐ability students are more likely to drop  out of school before reaching the level at which the test is implemented.  This implies that observed  average test scores in later grades likely overstate the quality of education in lower grades.   Adjusting  for this is difficult, as it requires some way to approximate the distribution of test scores among those  students who did not take the test.  However, a partial adjustment for this in the overall HCI comes  through the fact that in such cases, the enrollment rates used to calculate expected years of school will  also be lower, which in turn lowers the learning‐adjusted years of school measure.    A final caveat worth noting is that test scores are only an imperfect measure of learning.  There  are the usual concerns that test scores measure performance only on those items of the curriculum that  are covered in the test.  Beyond this, test scores also respond to other factors, including intrinsic  motivation on the part of test takers.  For example, extensive empirical evidence from schoolchildren in  Chicago suggests that small immediate financial rewards for good performance have substantial effects  on students’ standardized test scores (Levitt, List, Neckerman and Sadoff (2016)).  To the extent that  there are cross‐country differences in students’ intrinsic motivation when taking standardized tests, this  will be conflated with the quality interpretation of the test scores.  Student test scores also reflect the  influence of the home environment.  De Philippis and Rossi (2017) use US PISA data to study test scores  of children attending the same school, but whose parents immigrated from different countries.  Children  of parents who immigrated from countries with high test scores also tend to do better on test scores  themselves, holding constant the quality of education they received by focusing on children in the same  schools.  A2.4  Returns to Education    Values for the return to education parameter  can be anchored in the vast empirical literature  that estimates Mincer (1958)‐style regression of log wages on years of school.  Returns to education  naturally vary across levels of education, by gender, and across countries.  However, in the interests of  34    generating a simple and transparent index that focuses on the variation in the quantity and quality of  education across countries, the HCI uses a single benchmark value of  0.08, or 8 percent per year of  school, for all countries and all levels of school.  To put this value in perspective, consider the following  sets of estimates from the existing literature:   Montenegro and Patrinos (2014) estimate returns to education using household survey data  from 139 countries.  They find an overall average return to an additional year of school of 10.1  percent, and disaggregated returns of 10.6 (7.2) (15.2) for primary (secondary) (tertiary).   Averaging the primary and secondary returns which are most relevant for the HCI gives a value  of 8.9 percent.  Jedwab and Islam (2018) update these estimates in background analysis for the  2019 World Development Report, and find an average return to school of around 8 percent as  well.   In a highly‐influential review of the development accounting literature, Caselli (2005)  summarizes the empirical consensus on returns to school as 13 percent (for less than four  years), 10 percent (for four to eight years) and 7 percent (for more than eight years).  This  parameterization was also adopted in the Penn World Tables 9.0 estimates of human capital  (Feenstra, Inklaar and Timmer (2015)).  Assuming that primary and secondary school each last  six years, the baseline parameterization implies a return to primary school of  0.13 0.10 0.12 and a return to secondary school of  0.10 0.07 0.08.        The value of  0.08 is deliberately chosen to be at the low end of this range.  This is because  the vast majority of estimates of the return to school do not also control for health, while the human  capital measure in Equation (1) is intended to reflect the partial effects of education and health on  worker productivity.  To the extent that empirical studies of the return to education are unable to  control of health (or the factors determining health), the resulting estimates may overstate the partial  effect of education on productivity.  For more discussion of this point see Caselli (2014), who advocates  using a conservative estimate of the return to school in a similar setting.     A final issue concerns the fact that the vast majority of empirical estimates of the returns to  school use simply the number of years of education, and do not adjust for the quality of education.  This  may be a concern if the returns to learning‐adjusted years of school differ from returns to years of  school.  One study that addresses this point is Hanushek and Zhang (2009), who generate quality‐ 35    adjusted years of school using an approach similar to Evans and Yuan (2017), and then estimate  Mincerian returns to quality‐adjusted years of school in a sample of mostly OECD countries.  While the  country‐by‐country estimates of returns to quality‐adjusted school years differ somewhat from the  returns to unadjusted school years, on average they are fairly similar (Hanushek and Zhang (2009),  Figure 2).  This provides some justification for applying returns estimated using “raw” years of education  to expected learning‐adjusted years of school as is done here.        36    A3:  Health      A3.1  Basic Methodology and Data Considerations    The measure of health  in the description of the HCI in Section A1 should be interpreted as a  scalar index of “latent” health that summarizes the aspects of health that matter for worker  productivity.  Latent health cannot be observed directly, and so implementing the HCI requires a  mapping from unobserved “latent” health  to directly‐observable health indicator that serve as proxies,  as well as an estimate of the corresponding “return” to health, .  Weil (2007) proposes a strategy for  doing so by recognizing that wages as well as observable summary indicators of health status such as  adult height both respond to unobserved latent health.  In the case of wages, this could reflect channels  such as improved physical strength enabling greater work effort, as well as the effects of better health  on better cognitive skills, both of which then are rewarded with higher wages.15  A large literature has  also argued that trends in average adult height within a country can serve as a proxy for trends in the  overall health conditions in a country.16  Poor in utero and early childhood nutrition and health lead to  stunting among children, which in turn is reflected in shorter adult height, as well as a greater incidence  of poor health outcomes among adults.    A key advantage of adult height as an observable health indicator is that there are many micro‐ econometric estimates of the “return” to height obtained from extended Mincer regressions of log  wages on education, height and other controls.  Weil (2007) develops a latent variable representation of  the relationship between unobserved latent health, wages, and height, and demonstrates that this can  be used to replace unobserved latent health and its return,  , with observed height and its  estimated return,  , in the expressions for the health component of human capital in  Equation (1).   The interpretation in this case is not that height directly makes workers more productive.   Rather, the correct interpretation is that if latent health improves in such a way that height increases by  1 cm, then this will lead to an increase in worker productivity of   percent, i.e. height serves as an  observable proxy for unobservable latent health.                                                               15  In fact, Case and Paxson (2008) study data on test scores, height, and earnings in the US and the UK and  conclude that all of the effect of height on earnings operates through cognition in their sample.  16  Considering within‐country over‐time trends in height is crucial here, since cross‐country differences in average  stature also reflect cross‐country differences in genetic predisposition for height among different ethnicities.  37      A practical problem with this approach is that cross‐country data on adult height are relatively  scarce.  Moreover, the interpretation of the cross‐country variation is clouded by genetic differences in  populations of different countries.  To address this problem, Weil (2007) suggests using a more widely‐ available summary indicator of health, adult survival rates (ASR).  The basic insight is that ASR also  improves within countries over time with improvements in latent health, in the same way that adult  height does.  As a result, the within‐country over‐time relationship between improvements in adult  height and ASR can be used to transform the “return” to height into a “return” to ASR, and furthermore,  ASR can be more meaningfully compared across countries .17   Specifically, this means that   in the  expression for human capital can be replaced with  , where  ,   and  ,  is the slope coefficient from a regression of adult height on ASR.  Intuitively, this slope  coefficient captures how both adult height and adult survival rates improve when latent health  improves, and this relationship can be used to convert the “return” to height into a “return” to ASR.18   Again, however, the correct interpretation is not that there is a labour market “return” to adult survival  rates.  Rather, the correct interpretation is that when latent health improves to the extent that ASR  increases by one percentage point, then worker productivity increases by   percentage points.    A complementary strategy to solve the problem of data scarcity for adult height is to instead use  data on stunting in childhood, which has become increasingly available, particularly in low‐income  countries where stunting is common and recognized as an important marker of poor early childhood  development outcomes.  Although country coverage of stunting data is less complete than for ASR, a  benefit of stunting as a proxy for health is that there is direct evidence on the links between height in  childhood and adult height.  Evidence from cohort studies that track individuals over time provide  evidence that height deficits in childhood persist into adulthood.  This relationship can be used to create  a link between stunting rates and likely future adult height, which analogously is referred to as  , .  This can then be used to derive an alternative measure of the contribution of health  to future adult productivity,  , where  , .  The same caveats of interpretation apply to this measure as do to  .                                                                 17  Considering the within‐country over‐time variation controls for time‐invariant factors contributing to cross‐ country differences in height.  18  An underlying assumption here is that latent health is a single scalar index and that the observed indicators all  respond to this single underlying measure.  See Weil (2007), Section III.C for a discussion of the consequences of  relaxing this assumption.  38      The HCI uses stunting and ASR as two alternative observable proxies for the overall health  environment.  Absent a strong view on which of these is a better proxy, in countries where both are  available, a simple average of their contributions to productivity in the HCI is used, i.e.   is replaced  with  /2.  In the (mostly richer) countries where data on  stunting are not available, only    is used.  Since both   and    are measured in the same units, the unavailability of one or the other should not make a  systematic difference for the calculation of the contribution of health to productivity.      Both ASR and stunting rates are imperfect proxies for latent health.  The choice of these  measures is largely dictated by the scarcity of alternative widely‐accepted and broadly‐available cross‐ country data on non‐fatal health outcomes that could be used instead.  The main alternative would be  to use the measure of “years lived with disability” (YLD), that is constructed by the WHO and by the  Institute for Health Metrics and Evaluation (IHME) Global Burden of Disease project.  These estimates  draw on available survey‐based data on the prevalence of individual health conditions, that are then  aggregated across conditions using disability weights.  The measure of YLD, together with “years of life  lost” (YLL) based on age‐specific death rates due to various conditions, make up the well‐known  “disability‐adjusted life years” (DALY) measure.19       Although the calculation of YLD reflects the best efforts of the profession to piece together and  extrapolate the limited available direct information on non‐fatal health outcomes into a comprehensive  measure, it is not well‐suited for the HCI for three main reasons:   The first is that YLD nearly uncorrelated with per capita income across countries.  This is to some  extent by design.  As noted above YLD is based on measures of the prevalence of non‐fatal  health conditions.  As countries become richer and health care improves, conditions that                                                               19  To understand the mechanics of YLL, YLD, and DALY, consider this very stylized setting.  Individuals with   expected years of life remaining get sick with probability .  Conditional on being sick, there are two outcomes (i)  death, with probability , and (ii) survival, but with disability that results from having been sick, and that lasts for   years.  In this setting,   measures years of life lost:  i.e. the probability of getting sick and dying,  , times the remaining number of years of life lost, .  Similarly,  1  is the number  of years lived with disability, i.e. the probability of getting sick but surviving,    1 , times the number of  years the disability lasts,  , times a “disability weight”   between zero and one that measures the severity of the  disability.  Note that while both YLD and YLL decline when , the incidence or risk of getting sick falls, YLL and YLD  move in opposite directions when the mortality rate among the sick, , decreases.  In this case YLL falls but YLD  increases since now more survivors live with the disability that follows from having been sick.  Note that  1  unambiguously declines when both  and  decline (since   ).    39    previously were fatal become non‐fatal, and the survivors of these conditions are recorded as  living with any subsequent disability.  This makes it challenging to use this particular measure as  an indicator of health that leads to improvements in aggregate productivity.     The second is that data on YLD are extremely heavily imputed and extrapolated to compensate  for the scarcity of the primary sources measuring the relevant non‐fatal health conditions.   While such extrapolation is unavoidable if some estimate is required (no matter how  imprecise)20, it makes it difficult to use this measure in a policy advocacy index since the links  from the imputed health indicators to primary country data sources recognizeable by  policymakers in the country are complex.  In fact, in many cases, imputed estimates of specific  health conditions have no directly‐measured counterparts in the country itself. 21  In this case,  the policy advocacy role of the HCI can be undermined by relying on heavily‐imputed data, since  it risks “papering over” data gaps, reducing the incentives to fill these gaps with new data  collection work.   The third is that using imputed data makes it difficult to track and interpret changes over time in  the index.  To oversimplify, if health conditions are imputed using a measure of living standards  such as GDP per capita, then changes in GDP per capita (i.e. growth) will lead to changes in  imputed health indicators that may have little to do with actual health outcomes or policy  interventions to improve them.  While these limitations of more comprehensive but imputed indicators of non‐fatal health outcomes are  real, the limitations of the data on ASR and stunting used in the HCI are real as well.   Measurement of  ASR requires data on death rates by age.  While these are readily available in countries with strong vital  registries, in roughly the poorest quarter of countries such data are missing or incomplete.  In these  countries, the UN Population Division instead estimates death rates by age by linking together the  limited available age‐specific mortality data with “model life tables” capturing the typical pattern of  distribution of deaths by age.  As noted above, data on stunting are increasingly available, but primarily                                                               20  Commendably the GBD estimates of YLD are accompanied by confidence bands, which are very wide given the  underlying data limitations.  For example, for most countries the confidence bands for adult YLD encompass nearly  the entire range of point estimates of YLD, which transparently indicates the substantial uncertainty associated  with these estimates.  21  These gaps are illustrated using metadata reporting the dates of the primary sources underlying the GBD  estimates of the top 10 global causes of YLD.  Taking 2017 as a reference point, calculate the fraction of countries  for which no data source was available for each cause in the previous 10 years (so that the data necessarily were  imputed/extrapolated beyond 10 years).  This fraction is greater than 75 percent for developing countries, for 8  out of the top 10 global causes of YLD (anemia and diabetes are the only two exceptions).   40    for developing countries through household surveys such as the DHS, and typically are infrequently  collected.      The rest of Section 3 details the specific steps of this procedure for calibrating the effects of  health on worker productivity.  Section A3.2 discusses micro‐econometric estimates of the return to  height,  .  Section A3.3 discusses estimates of the relationship between adult height and adult  survival rates,  , .  Section A3.4 discusses how to calibrate estimates of the change in adult  height attributable to reductions in stunting,  , .    A3.2  Estimates of the Return to Height    Weil (2007) uses a baseline estimate of  0.034, i.e. one additional centimeter of  height raises earnings by 3.4 percent.  This evidence is taken from two previous studies comparing  height and earnings within twin pairs in the United States and in Norway.  The main strength of these  twin studies is their identification strategy of relying in random variations in birthweight between twins  as a plausibly exogenous source of variation in their eventual adult heights.  However, one might  reasonably be concerned with the external validity of these findings, since these studies are based on  pairs of twins in two advanced economies (Norway and the United States).  To assess this concern, it is  useful to briefly consider three other sets of estimates of returns to height from 19 other studies  covering a range of developed and developing countries.  All of these other estimates are based on  instrumental variables regressions of log wages on height, with instruments of varying degrees of  plausibility.  Conditional on the validity of the instruments, these should all recover the effect of height  on wages conditional on education, either because education is included in the regression, or because  the instrument is uncorrelated with omitted education (under the null hypothesis that the exclusion  restriction holds).    The first set can be found in Table 1 of Weil (2007),  which reports estimates of the return to  height conditional on education from three studies in Colombia, Ghana, and Brazil in the 8 to 9 percent  per year range.  The second set is summarized in Table 1 of Galasso and Wagstaff (2016), who  summarize five studies in developing countries, none of which overlap with those in the first set.  They  arrive at mean return to height of about 1.5 percent per year.  The third set of studies are summarized  in Horton and Steckel (2011).  Their Table 1 reports estimates from studies for 8 advanced economies  not covered in the previous two sets, with a mean return to height of 0.5 percent per centimeter.  Their  Table 2 reports studies for developing countries.  The three studies not included in the previous sets of  41    results provide returns to height ranging from 1.4 percent to 4.5 percent per centimeter.  In what  follows, the Weil (2007) preferred value of 3.4 percent is used as the baseline.   A reasonable range of  estimates has 6.8 percent as the upper bound (corresponding to the mean estimated return to height  across the 5 studies with estimates greater the Weil (2007) benchmark), and a lower bound of 1 percent  as the lower bound (corresponding to the mean estimated return to height in the remaining 13 studies  with estimated returns below 3.4 percent).22  A3.3  The Relationship Between Adult Height and Adult Survival Rates     The second key ingredient in the calculation is the estimated relationship between height and  adult survival,  , .  Weil (2007) estimates this using long‐run historical data on stature and  survival rates for 10 advanced economies over the 20th century, where there is considerable variation  within countries over time in adult height.  In his sample, average height varies from around 164 cm to  180 cm, and he obtains an estimate of  , 19.2.  To assess the robustness of this finding, the  same relationship is estimated using data on female height collected in 172 DHS surveys covering 65  developing countries  between 1991 and 2014.23   In this sample, female height exhibits comparable  variation to the historical dataset in Weil (2007), ranging from 148 cm to 163 cm.    In the roughly half of  the sample corresponding to non‐Sub‐Saharan African countries, a country‐fixed effects regression of  height on adult survival results in a slope coefficient of 19 and a standard error of 3.6, which is  extremely close to the Weil (2007) baseline estimate of 19.2.  In Sub‐Saharan Africa, the slope  coefficient is close to zero and very imprecisely estimated.  This likely reflects the large swings in adult  mortality rates due to the AIDS epidemic.  To assess this, the relationship is re‐estimated for all  countries, but excluding observations above the Sub‐Saharan Africa median for aids‐related death rates.   This gives a very similar estimate of 18.3 with a standard error of 3.4.   These estimates lie in the same  vicinity as the baseline estimate in Weil (2007).  This evidence suggests that the Weil (2007) estimate of  , 19.2 is reasonable to use in the baseline “return” to ASR of   , 0.034 19.2 0.65.                                                               22  All of the discussion here relies on the literature on the returns to height.  While less extensive, there are well‐ identified econometric estimates of the economic returns to a number of other measures of health, including iron  deficiency anemia, malaria, and hookworm infestation.  Subsequent updates of this background paper will  investigate how the estimates of overall health to productivity change when these other estimates of individual  health conditions and their returns are used in this framework.  23  The data come from the Health Equity and Financial Protection (HEFPI) Project at the World Bank.  Patrick  Eozenou and Adam Wagstaff kindly made this data available.  42        A3.4  The Relationship Between Stunting and Adult Height    An alternative approach to incorporating health into the human capital index is to use measures  of stunting in childhood directly as the observed proxy for latent health.  Stunting is measured as the  fraction of children under five years old whose height is more than two reference standard deviations  below the reference median, where the reference median and standard deviation are taken from WHO  standards for normal healthy child development.  Creating a link from stunting to the contribution of  latent health to productivity, requires evidence on the relationship between the proportion of children  who are stunted in childhood and average attained height of the population in adulthood.  Combining  this relationship with the estimated labour market returns to height creates a link from stunting in  childhood to worker productivity in adulthood operating through the channel of increased height.  This  subsection discussed two complementary approaches to obtaining this relationship    The first is a calibration based on a simple variant on the calculations and estimates in Galasso  and Wagstaff (2016).  They cite a number of cohort studies that provide evidence that having been  stunted as a child reduces attained adult height by approximately 6 centimeters.  Under the assumption  that average adult height conditional on stunting status in childhood does not change with the stunting  rate, they calculate the change in average adult height due to the elimination of stunting as this  difference of 6 centimeters multiplied by the fraction of the adult population that was stunted in  childhood.      This estimate may however be conservative because it assumes no change in the adult height of  children who were not initially stunted, even though these children are likely also to benefit from the  improvements in health that reduce the proportion of children who are stunted.  These wider effects  can be captured with an alternative calibration of how the mean of the distribution of adult height shifts  when childhood stunting falls.  Let   represent adult height and   represent the fraction of adults who  were stunted as children, i.e.  ≡  where   denotes height in childhood when stunting is  measured, and   represents the corresponding age‐specific height threshold for stunting in childhood.   Next consider three simplifying assumptions: (i) adult height is normally distributed, i.e. ~ , ; (ii)  the fraction of adults who were stunted as children is the same as the fraction of children who were  stunted when these adults were children, i.e.  ≡ , where  is the adult height threshold  43    corresponding to   in childhood; and (iii) the ordering of children by height in the under‐5 age group  where stunting is measured persists into adulthood.  Assumption (ii) enables the use of observed data  on stunting in childhood to measure the proportion of adults who were stunted as children, although  this requires abstracting from “catchup growth” as well as higher rates of mortality among stunted  children, both of which would lead to  .    Assumption (iii) ensures that the same group of  individuals who were stunted as children are also stunted as adults.  This assumption can be rationalized  by the high correlation between childhood and adult height.    As noted above, data on   is available, which by Assumption (ii) is equal to stunting in  adulthood, .  Estimates of the mean difference in adult height between adults who were not, and who  were, stunted as children,  ,  also exist and Assumption (iii) ensures that adults who were stunted as  children are also stunted as adults.  Together with Assumption (i) of normality, this implies two moment  conditions relating the data on  and   to the parameters of the distribution of adult height,  and :    (6)    | |   (7)  1   where  .  and  .  denote the normal distribution and density functions, and  (7) relies on the  properties of the truncated normal distribution.24      These two equations can be used to calibrate the changes in average adult height  associated  with reduced stunting rates .  One simple way for doing so is to use Equation (6) to trace out the  relationship between   and  for a fixed value of the standard deviation of height, .  Another way of  doing so is to solve Equations (6) and (7) to obtain this expression for mean adult height as a function of  the rate of stunting :                                                                  24  Specifically,  |  and  | .   44    1 (8)      This expression can be used to trace out the relationship between  and  for a fixed value of the mean  difference in height between adults who were and were not stunted as children,  .  Both of these  methods can be contrasted with the assumption in Galasso and Wagstaff (2016) in which the only effect  on adult average height comes through a reduction in the stunting rate, i.e.     Figure A3.4.1  Calibrated Relationship Between Adult Height and Stunting  166 Mean Adult Height in cm (µ) 164 162 160 158 156 154 152 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Fraction Stunted (q) Hol ding σ fi xed Holding d Fixed Galasso‐Wagstaff       Figure A3.4.1 plots the relationship between average adult height (on the vertical axis) and  stunting (on the horizontal axis) implied by these three methods.  To plot these graphs, set  156  centimeters, corresponding to the WHO height‐for‐age z‐score of ‐2 for 19 year‐olds (average across  male and female).  The thin blue line plots the relationship between mean adult height and stunting  holding fixed the standard deviation of height.25  The heavy black line shows the same relationship in                                                               25  The value at which  is held fixed matters for the calculation.  To make the series comparable, set  so that  Equations (6) and (7) hold at a stunting rate of  0.25 and the observed height difference of  6 in the study  by Victoria et. al. (2008) cited by Galasso and Wagstaff (2016).  Victoria et. al. (2008) report on young adult health  outcomes observed in the mid‐2000s in cohort studies that have tracked respondents since childhood.  Data from  the WHO‐UNICEF‐WB Joint Malnutrition Estimates databased indicate that stunting rates in the early 1990s (when  the respondents were children) in the five countries  were 19% (Brazil), 66% (Guatemala), 62% (India), 43%  (Philippines) and 32% (South Africa).  This range of values for the stunting rate is represented in the horizontal axis  of the figure.  In contrast, the value of  does not matter for the analysis since it only shifts the relationship  between stunting and adult height up and down, without changing the slope.  45    Equation (8), which holds fixed the height differential  , while the dashed red line shows the  relationship assumed in Galasso and Wagstaff (2016) which holds fixed mean height among adults who  were and were not stunted in childhood and varies only the proportion stunted.   Except at low rates of  stunting, the first two methods give a very similar relationship between mean adult height and stunting  rates in childhood.  Moreover, this relationship is steeper than the relationship assumed in Galasso and  Wagstaff (2016).  This is because their approach does not take into account the increases in height  among individuals who were not initially stunted as the stunting rate declines.  The slope of the dashed  red line is  6, while the average slope of the other two lines over the range where they coincide  is  10.2.  Consequently, a reduction in the rate of stunting  by ten percentage points raises attained  adult height by 10.2 0.1 or approximately one centimeter, or approximately 40 percent larger than in  the calibrations of Galasso and Wagstaff (2016).      The main advantage of this calculation is that it provides a very simple way to calibrate the  response of mean adult height to stunting in childhood, using only data on childhood stunting rates and  the estimate of the adult height differential from cohort studies.  An alternative approach to inferring  shifts in the mean of the distribution of height associated with reductions in stunting is to estimate them  directly.  This can be done using the same cross‐country panel of DHS surveys described in the previous  subsection.  These surveys contain data on the incidence of stunting, as well as average attained height  of children of different ages.  A country‐fixed‐effects regression of average height of two‐year‐olds on  the fraction of children who are stunted yields a slope coefficient of ‐0.12 and a standard error of 0.012.   This implies that a reduction in the stunting rate of 10 percentage points is associated with an increase  in average height among two‐year‐olds of 1.2 centimeters.  Under the assumption that height deficits in  two‐year‐olds persist into adulthood, this implies a reduction in average adult height of about the same  amount.  This estimate is slightly larger than but quite close to the one obtained by the calibration  approach discussed above.  To be conservative, the HCI uses the smaller of the two estimates by setting  , 10.2, with an overall “return” to reduced stunting of   , 0.034 10.2 0.35.        46    A4  The Human Capital Index  A4.1:  Putting the Pieces Together    This section draws together the discussion of the previous sections to summarize the overall  HCI, which is the product of three components:     (9)  ℎ ℎ     Using the notation from Equation (3), the three components of the index are formally defined as:  1 5 (10)  ≡   ∗ 1   ∗ ℎ ≡   (11)      ∗ / (12)  ℎ ≡       The baseline values for the returns to education and health are  0.08,   0.65 and  0.35 as discussed in the previous sections.  The probability of survival until age 5 is shown in  Figure 1.  The education component of the index is shown in Figure 5, and the health component of the  index is shown in Figure 6, separately for adult survival rates and stunting.   Expected learning‐adjusted  years of school range from around 3 years to close to 14 years in the best‐performing countries.  This  gap in expected learning‐adjusted years of school implies a gap in productivity relative to the benchmark  . of complete education of  0.4, i.e. the productivity of a future worker in  countries with the lowest expected years of learning‐adjusted school is only 40 percent of what it would  be under the benchmark of complete education.      For health, adult survival rates range from  60 to 95 percent, while the fraction of children not  stunted ranges from around 60 percent to over 95 percent.  Using ASR this implies a gap in productivity  . . . of  0.77, i.e. productivity of a future worker using the ASR‐based measure of  health is only 77% of what it would be under the benchmark of full health.  Using the fraction of children  . . . not stunted, this  implies a gap in productivity of  0.87, i.e. productivity  47    of a future worker using the stunting‐based measure of health is only 87% of what it would be under the  benchmark of full health.     The overall HCI is shown in Figure 7, and ranges from around 0.3 in the lowest countries to  around 0.9 in the highest countries.  This means that in countries with the lowest value of the human  capital index, the expected productivity as a future worker of a child born today is only 30 percent of  what it would be under the benchmark of complete education and full health.  A4.2:  Robustness To Alternative Weights    The calibrated returns to education and health, i.e.  ,  , and  , determine both the  range of the HCI as well as the relative weights on education and health in the HCI.  The higher are the  returns to education and health, the greater are the productivity differences implied by the differences  in learning‐adjusted school years and health.  In addition, higher (lower) values of the returns to health  relative to education place greater (lower) weight on the health component of the HCI.  To the extent  that countries have different relative positions in the education and health measures included in the  HCI, changing the relative weights on health and education can change countries’ relative positions in  the overall HCI.  However, these changes in relative positions are not very large because, not  surprisingly, the education and health measures included in the HCI are fairly highly correlated across  countries.    This can be seen in Figure 10, which shows the correlation between the baseline HCI reported in   Figure 7 and three alternative versions corresponding to three alternative estimates of the return to  height (which in turn feed into   and  ).  The baseline assumed return to an additional  centimeter of height is  0.034 or 3.4 percent.  As discussed in Section A3.2, a reasonable range  of values from the empirical literature goes from 1 percent to 6.8 percent.   Alternative versions of the  HCI using these estimates are shown in the top left and top right panels of Figure 10. They are  correlated with the baseline HCI at 0.99 in both cases.    Another way of assessing the robustness of the index to alternative weighting schemes is to  consider the (arbitrary) benchmark in which the education and health components of the index simply  are assumed to have equally‐large effects on worker productivity.  Specifically, let   and   denote  the largest and smallest observed values for learning‐adjusted years of school across countries, and  similarly let   and   denote the larges and smallest values of the health measure.  Then setting  48     corresponds to the assumption that moving from the bottom to the top of the  distribution of countries in health has the same effect on worker productivity as moving from the  bottom to the top of the distribution of education.  The range of observed outcomes for learning‐ adjusted years of school is about 11 years, while the range of observed outcomes for adult survival rates  is about 0.5, i.e.  22.  Using the baseline value of  0.08 and using  ,  implies   0.09 or 9 percent per centimeter (holding fixed  , 19.2),  which is much higher than is observed in the empirical literature.  An alternative version of the HCI using  this higher return to height, which in turn implies equal weights on education and health, is shown in  the bottom‐left panel of Figure 10.  Again, the correlation with the baseline HCI is very high at 0.99.   Overall this suggests that countries’ relative positions on the HCI are fairly robust to changes in the  calibrated returns to health and education that determine the relative weights on the components of  the HCI.  A4.3:  Gender Disaggregation    The components of the HCI, and therefore also the HCI itself, can be disaggregated by gender  for 126 countries.  Gender gaps are most pronounced for survival to age 5, adult survival, and stunting,  where girls on average do better than boys in nearly all countries.  Expected years of school is higher for  girls than for boys in about two‐thirds of countries, as are test scores.  The gender‐disaggregated overall  HCI is shown in Figure 9.  It is calculated by using the gender‐disaggregated components to evaluate the  overall HCI, while keeping the returns to health and education constant.  As a result, the gender  differences in this figure reflect only gender differences in the components of the HCI.  Overall, HCI  scores are higher for girls than for boys in the majority of countries.  The gap between boys and girls  tends to be smaller and even reversed among poorer countries, where gender‐disaggregated data also is  less widely available.  A4.4:  Uncertainty Intervals for the HCI and Its Components    All of the components of the HCI are measured with some error, and this imprecision naturally  has implications for the precision of the overall HCI.  This section briefly describes how imprecision in  the components of the HCI is measured, and the implications for imprecision in the overall HCI.  Formal  measures of imprecision are available for each of the components of the HCI, with the exception of  expected years of school, as follows:  49     Under‐5 mortality rates:  The UN Child Mortality Estimates program reports 90 percent  uncertainty intervals for under‐5 mortality rates.  These uncertainty intervals reflect imprecision  in the primary data sources (e.g. vital registries, household surveys, etc.) as well as imprecision  attributable to the smoothing mechanism that is used to generate annual estimates of these  rates.  For the median country in 2017, the 90 percent uncertainty interval is equal to 0.01 or a  range of 1 percentage point, while the median estimate of under‐5 mortality is 2 percent.  For  countries with higher estimated mortality rates, the uncertainty intervals can be larger:  for  example, the 75th  (90th) percentile of uncertainty intervals are 3.2 (5.3) percentage points wide.   Harmonized learning outcomes (HLOs):  As described above in Section A2.2, the calculation of  HLOs involves the application of a test x subject x grade‐specific conversion factor to the  country‐level average test score in its original units.  This means that there are two distinct  sources of uncertainty in the HLO calculation: (a) uncertainty around the country‐level average  scores in their original units, as reflected in the reported standard error around the country‐ level average , and (b) uncertainty in the calculation of the conversion factor.  The HLO database  quantifies the combination of these two sources of uncertainty through bootstrapping.   Specifically, 1000 random draws are taken from the distribution of the test x subject x grade‐ specific original score at the country level, assuming that the country‐level mean (across  students) score is normally distributed.  Then the HLOs are calculated using the 1000 samples of  original scores, and the 2.5th and 97.5th percentiles of the resulting bootstrapped HLOs are  reported as upper and lower bounds.  The HLOs used in the HCI are further aggregated to the  country‐year level as described in Section A2.2.  This aggregation is done using the reported HLO  estimates at the test x subject x grade level, and then repeated using the lower and upper  bounds of the test x subject x grade‐level scores.  The median HLO score as used in the HCI in  2017 is 424 TIMSS‐equivalent points, and the median range of the uncertainty interval is fairly  narrow at 12 points.  However, this range is larger for testing programs such as PASEC and  SACMEQ which have few “doubloon” observations on which the conversion factor is based, so  that uncertainty coming from variation in the conversion factor is larger.   Adult Survival Rates:  Adult survival rates (ASR) are compiled by the UN Population Division  using a similar process to the under‐5 mortality rates described above.  While there is  uncertainty in the primary estimates of mortality as well as the process for data modeling, UNPD  does not report uncertainty intervals.  Instead, uncertainty intervals produced in the IHME  Global Burden of Disease modelling process for ASR are used.  The point estimates for ASR in the  50    IHME and UNPD data are quite similar for most countries.  The ratio of the upper (lower) bound  to the point estimate of ASR in the IHME data is applied to the point estimate of ASR in the  UNPD data to obtain upper (lower) bounds on ASR.  The median uncertainty interval is 4.4  percentage points wide, while the median adult survival rate is 86 percent.  Uncertainty  intervals are substantially smaller (larger) for countries with higher (lower) ASR.  The 25th and  75th percentiles of the width of the uncertainty interval are 2.5 and 7.2 percentage points  respectively.   Stunting:  The UNICEF‐WHO‐World Bank Joint Malnutrition Estimates reports 95 percent  confidence intervals around estimates of stunting for about 40 percent of observations –  primarily those where the JME team has access to the record‐level survey data and can do  reanalysis.  These also correspond to the set of surveys for which gender‐disaggregated stunting  rates are available, and confidence intervals are reported for all gender‐disaggregated rates.  For  the median observation, the 95 percent confidence interval is just under four percentage points  wide.  Absent better alternatives, for the remaining observations in the JME database,  confidence intervals are imputed using the fitted values a regression of the width of the  confidence interval on the stunting rate. Looking at the cross‐section of most recently‐available  data for all countries in 2017, and after this imputation, the 95 percent confidence interval is 3.5  percentage points wide, while the median stunting rate is 22 percent.     Transforming the uncertainty intervals for the individual components of the HCI into uncertainty  intervals for the overall HCI is complicated by the fact that there is no information on the joint  distribution of uncertainty across components of the HCI.  To see why this matters, note that if  measurement error were uncorrelated across the different components, then the uncertainty intervals  for the overall HCI would be smaller than for the components since over‐estimates of some components  would be offset by under‐estimates of other components.  If by contrast measurement error were highly  correlated across components, then uncertainty intervals for the overall HCI would be larger than for  the individual components, as over‐estimates on one component would be compounded by over‐ estimates on other components, and vice versa.      Absent any information on the extent of correlation of measurement error across components,  the HCI uses the simple approach of constructing a lower (upper) bound of the uncertainty interval for  the overall HCI by assuming that each of the components is at its lower (upper) bound.  This approach is  conservative in the sense that it amounts to assuming that the measurement error is highly correlated  51    across components of the HCI.  On the other hand, these intervals understate the degree of uncertainty  around the overall HCI scores because they do not capture (a) uncertainty around the estimates of  expected years of school (for which uncertainty intervals are not available) and (b) uncertainty around  the estimates of the returns to education and health that are used to transform the components of the  HCI into contributions to productivity.    The resulting uncertainty intervals are shown in Figure 8, as vertical ranges around the value of  the HCI for each country.  The uncertainty intervals are moderate in size:  the median uncertainty  interval across all countries has a width of 0.03, while the HCI scores range from around 0.3 to 0.9.  For  some countries with less precise component data, the uncertainty intervals can be larger:  the 75th and  90th percentiles of the width of the uncertainty interval are 0.04 and 0.05 respectively.    Although crude, these uncertainty intervals are a useful way of indicating to users that the  values of the HCI for all countries are imprecise and subject to errors, reflecting the corresponding  imprecision in the components.  This should not be too surprising given the various limitations of the  component data described in previous sections.  The uncertainty intervals can also serve as an antidote  against the tendency to over‐interpret small differences between countries.  While the uncertainty  intervals constructed here do not have a rigorous statistical interpretation, they do signal that if for two  countries overlap substantially, the differences between their HCI values are not likely to be all that  practically meaningful.  This is intended to help to move the discussion away from small differences in  country ranks on the HCI, and towards more useful discussion around the level of the HCI itself and what  it implies for the productivity of future workers.            52    A5:  Linking the Human Capital Index To Future Income Levels and Growth  This section provides illustrative links from human capital to growth anchored in the logic of the  development accounting literature (see for example Caselli (2005) and Hsieh and Klenow (2011)).   It  follows much of this literature in adopting a simple Cobb‐Douglas form for the aggregate production  function:    (13)      where  is GDP per worker;   and   are the stocks of physical and human capital per worker; and  is  total factor productivity; and  is the output elasticity of physical capital.  When thinking about how  changes in human capital may affect income levels in the long run, it is useful to re‐write the production  function as follows:  (14)      In this formulation, GDP per worker is proportional to the human capital stock per worker, holding  constant the level of total factor productivity and the ratio of physical capital to output,  .  This  formulation can be used to answer the following question:  “By how much does an increase in human  capital raise output per worker, in the long run after taking into account the increases physical capital  that is likely to be induced by the increase in human capital?”.  The answer to the question is that output  per worker increases equiproportionately to human capital per worker, i.e. a doubling of human capital  per worker will also lead to a doubling of output per worker in the long run.    Linking this framework to the human capital index requires a few further steps.  First, following  much of the existing literature, assume that the stock of human capital per worker that enters the  production function,  , is equal to the human capital of the average worker.26 Second, the human                                                               26  This is by no means innocuous, because it embodies the strong assumption that workers with different levels of  human capital are perfectly substitutable after taking into account their individual productivity differences.  To  take a highly simplified and memorable example (due to David Weil) of where perfect substitutability breaks down,  note that although the educational human capital of four unskilled workers probably is lower than that of one PhD,    53    capital of the next generation, as measured in the HCI, and the human capital stock that enters the  production function, need to be linked.  This can be done by considering the scenarios outlined in the  main text.  Imagine first a “status quo” scenario in which the expected learning‐adjusted years of school  and health as measured in the HCI today persist into the future.  Over time, new entrants to the  workforce with “status quo” health and education will replace current members of the workforce, until  eventually the entire workforce of the future has the expected learning‐adjusted years of school and  level of health captured in the current human capital index.  Let  ,  denote the future  human capital stock in this baseline scenario.  Contrast this with a scenario which the entire future  workforce benefits from complete education and enjoys full health, resulting in a higher human capital  ∗ ∗ stock  ∗ .        It is possible to compare eventual steady‐state GDP per worker levels in the two scenarios using  Equation (14), assuming that levels of TFP and the physical capital‐to‐output ratio are the same in the  two scenarios, to obtain:  , ∗ ∗ (15)  ∗   ∗   This expression is the same as the human capital index in Equation (3), except for the term  corresponding to survival to age 5 (since children who do not survive do not become part of the future  workforce).  This creates a close link between the human capital index and growth.  Disregarding the  (small) contribution of the survival probability to the HCI, Equation (15) says that a country with an HCI  equal to   could have future GDP per worker that would be 1/  times higher in the future if its citizens  enjoyed complete education and full health (corresponding to  1).  For example, a country such as  Morocco with a HCI value of aroudn 0.5 could in the long run have future GDP per worker in this  scenario of complete education and full health that is  2  times higher than in the status quo  . scenario.  What this means in terms of average annual growth rates of course depends on how “long”  the long run is.  For example, under the assumption that it takes 50 years for these scenarios to                                                               the four unskilled workers are undoubtedly more productive when it comes to moving a piano.  See Jones (2014)  for alternative human capital aggregators that relax the assumption of perfect substitutability.  Jones (2014) argues  that allowing for complementarities between workers of different skill levels substantially increases the role of  human capital in accounting for cross‐country income differences.  However, Caselli and Ciccone (2017) point out  that this interpretation ignores the important role of cross‐country differences in productivity in driving the skill  premia that in turn drive the conclusions in Jones (2014).  54    materialize, then a doubling of future per capita income relative to the status quo corresponds to  roughly 1.4 percentage points of additional growth per year.    The calibrated relationship between the HCI and future income levels described here is simple  because it focuses only on steady‐state comparisons.  In related work, Collin and Weil (2018) elaborate  on this by developing a calibrated growth model that traces out the dynamics of adjustment to the  steady state.  They use this model to trace out trajectories for per capita GDP and for poverty measures  for individual countries and global aggregates, under alternative assumptions for the future path of  human capital.  They also calculate “equivalent” increases in investment rates in physical capital that  would be required to deliver the same increases in output associated with improvements in human  capital.    References  Altinok, Nadir, Noam Angrist, and Harry Patrinos (2018).  “Global Dataset on Education Quality 1965‐ 2015”.  World Bank Policy Research Working Paper No. 8314.  Ashraf, Quamrul, Ashley Lester and David Weil (2009).  “When Does Improving Health Raise GDP?”, in  Daron Acemoglu, Kenneth Rogoff and Michael Woodford, eds. NBER Macroeconomics Annual.  Case, Ann and Christina Paxson (2008).  “Stature and Status:  Height, Ability, and Labour Market  Outcomes”.  Journal of Political Economy. 116(3): 499‐532.  Caselli, Francesco (2005).  “Development Accounting”.  Handbook of Economic Growth.  Caselli, Francesco (2014).  “The Latin American Efficiency Gap”.  Manuscript, London School of  Economics.  Caselli, Francesco and Antonio Ciccone (2017).  “The Human Capital Stock:  A Generalized Approach –  Comment”.  Manuscript, London School of Economics.  Collin, Matthew, and David N. Weil. (2018). “The Effect of Increasing Human Capital Investment on  Economic Growth and Poverty: A Simulation Exercise.” World Bank Policy Research Working  Paper No. 8590.  De Philippis, Marta and Federico Rossi (2017).  “Parents, Schools, and Human Capital Differences Across  Countries”.  Manuscript, Johns Hopkins University.  Evans, David and Fei Yuan (2017).  “The Economic Return to Interventions That Increase Learning”.   Background Paper for World Development Report (2018).   https://www.riseprogramme.org/sites/www.riseprogramme.org/files/inline‐ files/Evans%20Returns%20to%20Learning%20‐%20v2017‐06‐09%20CLEAN.pdf.   55    Feenstra, Robert C., Robert Inklaar and Marcel P. Timmer (2015), "The Next Generation of the Penn  World Table" American Economic Review, 105(10), 3150‐3182, available for download at  www.ggdc.net/pwt.   Filmer, Deon, Halsey Rogers, Noam Angrist,  and Shwetlena Sabarwal (2018).  “Learning‐Adjusted Years  of Schooling (LAYS): Defining a New Macro Measure of Education.”  World Bank Policy Research  Working Paper No. 8591  Gakidou, Emmanuela, Krycia Cowling, Rafael Lozano and Chris Murray (2010).  “Increased educational  attainment and its effect on child mortality in 175 countries between 1970 and 2009: a  systematic analysis”.  The Lancet. 376(9745):18‐24.  Galasso, Emanuela and Adam Wagstaff (2016).  “The Economic Costs of Stunting and How to Reduce  Them”.  World Bank Policy Research Note 17/5.  Gao, Qiufeng, Yaojiang Shi, Hongmei Yi, Cody Abbey, and Scott Rozelle (2017).  “Reading Achievement in  China’s Rural Primary Schools: A Study of Three Provinces”.  Stanford University, Freeman Spogli  Institute Working Paper, available at https://fsi.stanford.edu/publication/reading‐achievement‐ chinas‐rural‐primary‐schools‐study‐three‐province.  Hall, Robert and Chad Jones (1999).  “Why Do Some Countries Produce So Much More Output Per  Worker Than Others?”.  Quarterly Journal of Economics.  114(1):83‐116.  Hanushek, Eric and Lei Zhang (2009). "Quality‐Consistent Estimates of International Schooling and Skill  Gradients." Journal of Human Capital 3(2): 107‐143.  Hanushek, Eric, Jens Ruhose, and Ludger Woessmann (2015).  “Human Capital Quality and Aggregate  Income Differences:  Development Accounting for US States”.  Hoover Institutions Economics  Working Paper No. 15112.  Horton, Sue and Richard Steckel (2011).  Copenhagen Consensus Assessment Paper on Malnutrition.    Hsieh, Chang‐Tai and Peter Klenow (2010).  “Development Accounting”.  American Economic Journal  (Macroeconomics).  2(1):207‐223.  Jedwab, Remi and Asif Mohammed (2018).  “Returns to Education and Experience”.  Background Paper  for 2019 World Development Report, in progress.  Jones, Benjamin (2014).  “The Human Capital Stock:  A Generalized Approach”.  American Economic  Review.  104(11):3752‐3777.  Jorgenson, Dale W., and Barbara M. Fraumeni (1989). “The Accumulation of Human and Nonhuman  Capital, 1948‐1984. In Robert E. Lipsey and Helen S. Tice, eds. The Measurement of Savings,  Investment and Wealth Studies in Income and Wealth. Chicago: University of Chicago Press.  Kaarsen, Nicolai (2014).  “Cross‐Country Differences in the Quality of Schooling”.  Journal of  Development Economics. 107: 215‐224.  Klenow, Peter and Andres Rodriguez‐Clare (1997).  “The Neoclassical Revolution in Growth Economics:   Has It Gone Too Far?”.  NBER Macroeconomics Annual.  56    Lagakos, David, Benjamin Moll, Tommaso Porzio, Nancy Qian and Todd Schoellman (2018).  “Life Cycle  Wage Growth Across Countries”.  Journal of Political Economy, forthcoming.  Levitt, Steven, John List, Susanne Neckerman and Sally Nadoff (2016).  “The Behavioralist Goes to  School: Leveraging Behavioral Economics to Improve Educational Performance.”  American  Economic Journal:  Economic Policy.  8(4): 183‐219.  Manuelli, Rodolfo and Ananth Seshadri (2014).  “Human Capital and the Wealth of Nations”.  American  Economic Review.  104(9):2736‐2762  Mincer, Jacob (1958).  “Investment in Human Capital and Personal Income Distribution”.  Journal of  Political Economy.  66(4):281‐302.  Montenegro, Claudio and Harry Patrinos (2014). “Comparable Estimates of Returns to Schooling Around  the World”. World Bank Policy Research Working Paper No. 7020.   OECD (2016).  “PISA 2015 Results (Volume 1).  Excellence and Equity in Education”.  Paris, OECD.   http://dx.doi.org/10.1787/9789264266490‐en.   Patrinos, Harry Anthony, and Noam Angrist. 2018. “A Global Dataset on Education Quality: A Review and  an Update (1965–2018).”  World Bank Policy Research Working Paper No. 8592.  Ravallion, Martin (2011).  “Mashup Indices of Development”.  World Bank Research Observer.  27:1‐32.  Schoellmann, Todd (2012).  “Education Quality and Development Accounting”.  Review of Economic  Studies.  79:388‐417.  UN (2016).  “Guide on Measuring Human Capital” (https://www.unece.org/index.php?id=44704).    UNESCO, World Bank, UNICEF and Global Partnership for Education (2014).  “Education Sector Analysis  Methodological Guidelines”. (https://www.globalpartnership.org/content/methodological‐ guidelines‐education‐sector‐analysis‐volume‐1).      United Nations University (2012).  “Inclusive Wealth Report”.  Cambridge:  Cambridge University Press.  Victora,  C. G., et al. (2008). “Maternal and Child Undernutrition: Consequences for Adult Health and  Human Capital. The Lancet. 371(9609):340‐57.  Weil, David (2007).  “Accounting for the Effect of Health on Economic Growth”.  Quarterly Journal of  Economics.  156‐1306.  World Bank (2018).  “The Changing Wealth of Nations”.     57