A Global Count of the Extreme Poor in 2012: Data Issues, Methodology and Initial Results

The 2014 release of a new set of purchasing power parity conversion factors (PPPs) for 2011 has prompted a revision of the international poverty line. In order to preserve the integrity of the goalposts for international targets such as the Sustainable Development Goals and the World Bank's twin goals, the new poverty line was chosen so as to preserve the definition and real purchasing power of the earlier $1.25 line (in 2005 PPPs) in poor countries. Using the new 2011 PPPs, the new line equals $1.90 per person per day. The higher value of the line in US dollars reflects the fact that the new PPPs yield a relatively lower purchasing power of that currency vis-à-vis those of most poor countries. Because the line was designed to preserve real purchasing power in poor countries, the revisions lead to relatively small changes in global poverty incidence: from 14.5 percent in the old method to 14.1 percent in the new method for 2011.In 2012, the new reference year for the global count, we find 12.7 percent of the world's population, or 897 million people, are living in extreme poverty. There are changes in the regional composition of poverty, but they are also relatively small. This paper documents the detailed methodological decisions taken in the process of updating both the poverty line and the consumption and income distributions at the country level, including issues of inter-temporal and spatial price adjustments. It also describes various caveats, limitations, perils and pitfalls of the approach taken.


Introduction
The estimated number of people in the world that live in extreme poverty has become an increasingly important (inverse) indicator for measuring development progress. The first Millennium Development Goal (MDG) aimed to halve the share of people living in extreme poverty between 1990 and 2015. 1 In 2013, the World Bank adopted a goal of ending extreme poverty by 2030. 2 In 2015, the United Nations adopted a goal of 'ending poverty in all its forms' by 2030 among the Sustainable Development Goals (SDGs). 3 Several national governments, bilateral development agencies and non-governmental organizations are also focusing their efforts on reducing, and ultimately ending, extreme poverty. As one of the most prominent indicators of economic development, the level and trends in extreme poverty are therefore a topic of great interest to both policymakers and the public at large.
Despite the strong policy focus on ending extreme poverty globally, its definition and measurement remain challenging endeavors, subject to much debate regarding the most appropriate concepts, methods and data. At the core of this debate is how to compare the standards of living of widely different peoples, consuming vastly different goods and services, all priced in different currencies. One dimension of this question is how to define a common threshold (or poverty line) across countries and over time, which represents the same standard of living, below which a person is considered poor.
Since the early 1990s, the World Bank has monitored global extreme poverty using an international poverty line that was explicitly based upon the national poverty lines of some of the poorest countries in the world. For ease of communication, this international poverty line has always been expressed in U.S. dollars (USD), but when used for measuring poverty, the line is converted into local currencies through purchasing power parity (PPP) exchange rates, in an 1 Target 1.A of the United Nations Millennium Development Goals (MDGs) is to "halve, between 1990 and 2015, the proportion of people whose income is less than $1.25 a day." For more details on the MDGs, see www.un.org/millenniumgoals/. 2 In April 2013, the World Bank Group announced the goal of ending extreme poverty, which it defined specifically as reducing "the percentage of people living with less than $1.25 a day […] to no more than 3 percent globally by 2030." For more details, see: www.worldbank.org/content/dam/Worldbank/document/WB-goals2013.pdf. 3 attempt to ensure that it has the same purchasing power in every country. The specific PPP conversion factors used for this exercise are those for private consumption, which come from the International Comparison Program (ICP). 4 Over the last few decades, the ICP has at various times collected price data on a large number of goods and services and in a large number of countries, with each assessment providing new information on how price levels compare across countries. Each release of new PPP data has led both to revisions of the international poverty line, and to re-assessments of the relative differences in wellbeing across countries and regions. The first international poverty line that was based on a sample of national poverty lines was set at $1.01 using 1985 PPPs, by Ravallion, Datt and van de Walle (RDW, 1991) and used in the 1990 World Development Report (World Bank, 1990). Chen and Ravallion (2001) later updated this to $1.08 per day, using the 1993 PPPs. With the release of the 2005 PPPs and a new set of national poverty lines, Ravallion, Chen and Sangraula (RCS, 2009) proposed a new poverty line of $1.25 per day, which is the line that has been used by the World Bank to measure poverty since 2009. Each of these updates has also been associated with revisions in the number and geographical composition of global poverty, often in ways that can be jarring and cast into doubt the reliability of the global poverty estimates. 5 The 2011 PPP conversion factors, released by the ICP in 2014, once again led to substantial revisions in relative price levels, the magnitude of which varies widely across countries.
Inevitably, incorporating the new information on relative price levels that is contained in these conversion factors has implications for the measurement of global poverty and, more specifically, for the definition of the international poverty line. In considering how best to update the line to take account of this new information, we have been guided by three central principles.
First, international comparisons of living standards should be based on the most recent reliable information on relative price levels, unless there is sufficient reason to suspect that it is less accurate than previous estimates. Second, we have sought to minimize changes to the real value of a poverty line that has become an important component of a system of international goals and targets, so as to avoid "shifting the goalposts". We have therefore sought to keep the definition of the line unchanged, and its new value as close as possible to that of the $1.25 line (in 2005 PPPs) 4 in real terms. Third, when defining "real terms", we have focused on the price levels that matter most for measuring global poverty, namely those faced by the world's poorest countries.
This paper provides a new set of global poverty estimates from 1990 to 2012, using the 2011 PPPs and an international poverty line that was revised in accordance with these three principles. The new line is still derived from the same fifteen national poverty lines that Ravallion et al. (2009)  The estimates produced in this paper apply the same methods as Chen and Ravallion (2010) and are based on data from PovcalNet, the database maintained by the World Bank for monitoring global poverty in recent years. 7 The most recent version of PovcalNet, contains data from more than one thousand surveys covering 131 developing countries, and 21 high income countries. In addition to household surveys and PPP data, the poverty estimates rely on intertemporal and spatial price deflators and adjustments, which can have large effects on both the international poverty line and on estimates of poverty rates for individual countries. The paper 6 A second poverty line of $3.10 in 2011 PPPs has also been proposed as the comparable equivalent to the $2 dollar-a-day poverty line in 2005 PPPs, commonly used as a poverty line for middle-income countries. 7 PovcalNet is perhaps the most commonly used data tool for estimating global poverty counts. It is an online tool, maintained by the World Bank, which allows analysts to specify parameter values such as the global poverty line, and then estimate the number of poor people in the world based on their assumptions. For more details, see: http://iresearch.worldbank.org/PovcalNet/index.htm. 5 also documents the methods used for aggregating poverty globally and the alignment of estimates to a particular reference year.
The remainder of the paper is structured as follows. The next section briefly reviews the literature on global poverty measures, with a particular emphasis on the use of PPP conversion factors and the impact they have had on international poverty measures in the past. Section 3 describes in detail the data and methods used in the latest revision of the global poverty numbers.
The fourth section describes how the $1.25 line has been updated to $1.90 based on the 2011 PPPs, and the final section presents the 2012 count and reports on the revised global poverty numbers from 1990 to 2012.

A brief history of global poverty measurement using PPPs
The precise methods used for measuring global poverty have changed over time, but one guiding principle has endured, namely that the global extreme poverty line should reflect how the world's poorest countries estimate a minimum threshold of living which meets basic needs in their societies. One instrument that has consistently been used for assessing minimum basic needs has been national poverty lines. Absolute national poverty lines, when well constructed, are anchored in core biological (e.g. caloric) needs, but also reflect country context and values, and often are robustly debated by the press and citizens. This section describes each of the major revisions to the international poverty line, which are also synthesized in Table 1.
One of the first estimates of internationally comparable poverty measures was conducted by Ahluwalia, Carter and Chenery (1979) who use India's national poverty line (46 th percentile of the per capita income distribution) to estimate the world's poverty incidence, using the 1975 PPPs from the ICP. Their estimate was based on consumption and income data from 25 countries. While the global coverage of this estimate was limited, 8 this was the first attempt to measure global poverty with a common absolute poverty line. It also began the practice of measuring international poverty based on national lines, converted at PPPs.
Since the 1990s, the international poverty lines used by the World Bank have been defined on the basis of information from a sample of the national poverty lines used in some of the world's poorest countries. Ravallion, Datt and van de Walle (RDW, 1991) examine 33 6 national poverty lines and note that six countries among the poorest in this sample (Bangladesh, Indonesia, Kenya, Morocco, Nepal and Tanzania) were all within one dollar of a poverty line of USD 31 per person per month at 1985 PPPs. The similarities in the value of the national poverty line for these six countries then served as the basis for the original "dollar-a-day" global poverty line. 9 The resulting global count of the poor for 1985 was 1.1 billion persons based on data for 22 countries (with extrapolation models for an additional 64 countries).
With new household-survey data and expanding country coverage, global poverty estimates were regularly updated, with some substantial modifications over time. These modifications typically occurred in response to each new ICP price data collection exercise, and the subsequent release of new PPP exchange rates, or conversion factors. The first major revision to the dollar-a-day line came when Chen and Ravallion (2001) updated the line based on the then more recent 1993 PPP data. They re-estimated the global poverty line following the same basic idea of basing the global poverty line on a sub-sample of national poverty lines, but used the median national poverty line from the ten lowest national poverty lines from the RDW sample of national poverty lines. This resulted in a revised global poverty line of $1.08 a day in 1993 PPP prices.
After the 1993 PPP data, the ICP did not release a significant update to the PPP data until the 2005 benchmark year. In part due to the very long gap in data updates, the release of the 2005 PPP data had significant implications for both the global poverty line and global poverty estimates. In preparing an update that incorporated this change in PPPs, Ravallion, Chen and Sangraula (2009) also compiled a new sample of national poverty lines from 74 countries and used it to re-estimate the global poverty line. They observed a positive association between the value of the poverty line and national per capita consumption for the large majority of countries, but found that the relationship was flat for the fifteen poorest countries in their sample. 10 The 7 simple average of the national poverty lines for these fifteen countries was $1.25 in 2005 PPP terms and this then became the revised global poverty line.
[INSERT FIGURE 1 APPROXIMATELY HERE] Using this new poverty line and the 2005 PPPs, Chen and Ravallion (2010) found that the number of poor was much larger than had previously been thought. The $1.08 poverty line and 1993 PPPs suggested that 1.3 billion people were poor all over the world. However, the new This large revision in the estimated count of poor people coincided with a growing focus on the global poverty measures in the international development community, and brought about a vigorous debate on how best to measure global poverty. As one example, Deaton (2010) presents the regional profile of global poverty in 1993 when assessed using the 1985, 1993, and 2005 PPPs and the matching global poverty lines for each benchmark year. He makes a couple of points: First, there is substantial variation in both the regional profile and overall level of poverty depending on which PPP and poverty line are used. Second, the $1.01 and $1.08 poverty lines combined with their benchmark PPP data produced very similar global poverty counts for 1993.
Deaton essentially argues that the $1.25 line and the 2005 PPPs resulted in a higher estimated poverty count primarily because of an upward revision of the value of the international poverty line in poor countries, rather than due to the effect of the 2005 PPPs.
A more general point that arises from Deaton's critique -and which the 2011 revisions we are currently incorporating reinforce -is that global poverty estimates have generally been -8 and remain -very sensitive to the introduction of new PPP conversion factors. As the ICP changes the way in which it collects and combines price data from different countries in an attempt to construct the best possible multilateral price index, estimates of the relative costs of living across nations can change substantially. To the extent that PPPs are used to compare living standards in poverty measurement, these changes inevitably -and occasionally powerfullyaffect poverty estimates.

The 2011 PPPs and the global poverty estimates 12
In 2014, the ICP released the 2011 PPPs, reflecting some key changes in coverage and methodology. Country coverage of the ICP increased from 146 economies in the 2005 round to 199 economies in the 2011 round. The 2011 PPPs are based on data from these 199 economies, which represent 97 percent of the world's population and approximately 99 percent of the world nominal GDP (in U.S. dollars using exchange rates). In addition to improved coverage globally, the 2011 PPPs also increased coverage within countries. In particular, World Bank (2015a) reports that the 2011 ICP saw improved coverage of China. The 2005 PPP data for China are based on collecting data from 11 cities (and surrounding areas), but the 2011 PPP data come from nationwide surveys covering urban and rural areas in all provinces. In terms of methodological changes, the ICP 2011 aimed to improve linkages across regions of the world by introducing a "global core list" of 618 items for all countries to include. 13 Once again, potentially due to these changes in ICP methodology, there is scope for the new price data to alter the overall profile of global poverty. The 2011 PPPs significantly revise relative price levels, particularly between rich and poor countries. In broad strokes, the main difference is that relative prices are substantially lower in developing countries according to the 2011 ICP round. As a result, incomes in PPP-adjusted dollars are correspondingly higher in poor countries than previously indicated by the 2005 PPPs. This has fostered a new debate about the relative merits and quality of these two ICP rounds. Deaton and Aten (2014) argue that the 2011 PPPs are superior to the 2005 PPPs, and in part reverse a bias in the 2005 PPP estimates. The 2005 PPPs are essentially the product of two 9 price indices -one index established purchasing power parity within regions and the other compared across regions. The cross-region index is based on data from a set of 18 so-called "ring" countries in which a distinct commodity list was used to price out goods that were not unique to any particular region. Deaton and Aten argue that the cross-region ring index is the key source of error in the 2005 PPP data, resulting in an overestimation of the relative price levels in Africa, Asia and Western Asia by 20 to 30 percent. Along similar lines, Inklaar and Rao (2014) re-construct the 2005 PPP index following a similar methodology that was used for the 2011 PPP index and find that this methodological difference in the construction of the PPP factors explains a significant portion of the difference in relative price levels.
Findings from Ravallion (2014) lead to almost the opposite interpretation, suggesting that there are potentially significant concerns about the 2011 PPP data. He argues that the downward drift in prices observed for much of Asia (but not China) is in contrast to what would be expected given the observed rate of economic growth. A part of his interpretation of the data rests on the so-called "dynamic Penn Effect" (Ravallion, 2013), which suggests that the ratio of the PPP index to the market exchange rate rises with economic growth. Ravallion offers a hypothesis that over time the bundle of goods used for the PPP index has become more heavily weighted towards internationally traded goods (for which prices exist) and this has led to a downward shift in price levels relative to market exchange rates (conditioning on growth rates).
Bypassing the debate regarding the quality of the 2011 PPP data, others have carried out quick analyses on how the 2011 PPP data would affect global poverty counts. Immediately after the release of the 2011 PPP data, Dykstra, Kenny and Sandefur (2014) estimated that in 2010, the share of people in the developing world living below the $1.25 per day poverty line declined from 19.7 percent to 8.9 percent "overnight". Their approach was to adjust the $1.25 poverty line, which is denominated in 2005 USD, to 2011 by using the U.S. inflation rate. This adjustment brings their poverty line to a value of $1.44 in 2011 USD. This methodology differs significantly from the approach followed by RCS and described in Jolliffe et al. (2014) for setting an international poverty line that reflect the poverty lines of some of the poorest countries in the world. Chandy and Kharas (2014) follow an approach similar in spirit (though different in the details) to the methodology used in RCS and re-estimate the poverty line accounting for changes in local price levels. They provide several candidate estimates of the count of the poor, 10 but hone in on an estimate where the number of poor is about 300 million lower in 2010 as a result of using the 2011 PPP data (rather than the 2005 PPP data). Jolliffe and Prydz (2015) note that all of the early estimates of global poverty based on the 2011 PPPs had been based on leaving in place an adjustment factor for presumed urban bias in India, China and Indonesia. The parameterization of the adjustment factor was specific to details of the fieldwork of the 2005 PPPs, and needed to be either dropped from the analysis or re-calibrated for the re-design of the 2011 PPP fieldwork. They show that the goal of reducing extreme global poverty to less than 3 percent by 2030 remained essentially just as ambitious with the 2011 PPPs as with the 2005 PPPs.

Data and methodology
A number of choices regarding data sources and methods are made in the construction of the global poverty estimates. These decisions reflect a mix of theoretical foundations, empirical analysis, and knowledge of country and region context. The present update of global poverty numbers (for the year 2012) follows closely the methods used by Chen and Ravallion (2010) based on the 2005 PPPs and PovcalNet data. While these estimates have been widely used for the monitoring of the MDGs and World Bank goals, some of the details that underpin them have not always been well understood. This section aims to provide an extensive description of the data, assumptions and methods used, and to highlight their merits and limitations. 14

Individual-level welfare measures from household surveys
At the core of the World Bank's global poverty measurement is a growing collection of income and consumption distributions, primarily derived from national household survey data.
For the present global poverty estimates, the World Bank uses income and consumption distributions from more than one thousand surveys covering 131 developing countries. The estimates are based on survey interviews of more than 2 million households, representative of 95 percent of the population in the developing world. While there has been a dramatic improvement in the availability of consumption and income data over the past two decades, this improvement in the stock has been accompanied by widening heterogeneity of the surveys in terms of instruments (for example, questionnaires) and methods used by countries to measure household consumption, expenditures and income. This widening heterogeneity poses an important challenge to the fundamental assumption in measuring global poverty, that the underlying distributions of wellbeing (whether income or consumption) are comparable across countries (and within countries, over time).
There are several reasons why this assumption of comparability is challenging, but one key reason is that the surveys underlying the welfare distributions used in PovcalNet are carried out by National Statistical Offices (NSOs) as part of national efforts to monitor poverty and other aspects of national wellbeing. The surveys are not designed for the purpose of international comparisons and global monitoring; they are designed to serve the specific needs and interests of each particular country. Different countries use different concepts, methods and questionnaire designs for estimating household consumption and income, and they frequently change the questionnaires over time. This heterogeneity limits the comparability of estimates between countries, and sometimes also within countries over time.
One issue of comparability is linked to the fact that essentially all countries in the PovcalNet database measure wellbeing (either in the form of consumption or income) at the level of the household. So one decision that needs to be made is how to convert a household level measure into an individual level measure. For many countries, making the wellbeing aggregate comparable across individuals means adjusting for differences in needs between different ages of people, typically based on adult-equivalence scales. It may also mean adjusting for household economies of scale, which aims to account for the possibility that a household of four people may need less than twice what a household of two needs to reach the same level of wellbeing.
The basic idea is that there may be some important goods, such as housing, that can be consumed by multiple individuals at the same time at a very low additional cost (ie. non-rivalrous goods). 15 The key issue for global poverty measurement is that, strictly speaking, a country that allocates household consumption to individuals within the household on a per capita basis has a consumption measure that is not comparable with a country that allocates consumption to 12 individuals based on adult-equivalence. 16 To treat all countries the same, all consumption and income measures in PovcalNet are in per capita terms. 17 Another important issue of comparability is the mix of income and consumption data in PovcalNet. Table 2 reveals that approximately 75 percent of the countries in the PovcalNet database have data on per capita consumption, while the remaining countries -mostly in Latin America and the Caribbean -have income per capita. Income and consumption measures are neither conceptually nor empirically comparable measures of welfare. Conceptually, income is usually described as defining the opportunity set, while consumption defines realized outcomes.
While incomes are generally thought to be more volatile, consumption is expected to somewhat smooth out these fluctuations (assuming concave utility functions). These are all reasons why it is expected that consumption and income will differ, and treating these as comparable measures of wellbeing is a potential concern.
Empirically, these differences manifest themselves in two main ways. The first is that, by definition, zero income is a feasible value, while zero consumption is not a feasible value (someone with zero consumption cannot exist). Indeed this difference is observable in the datathere is essentially no mass point in any country with zero consumption, 18 but many countries that use income data have a significant mass of zero incomes in the data, all of which are treated as being poor. Latin America, in particular, predominantly uses income to measure poverty, and in many countries there are at least a few percent of the observations that are zero.
The difference between consumption and income can also be due to issues of data quality. It is typically presumed that data quality differs between income and consumption, often based on the size and formality of the economy. In very poor countries, with little formal 16 Consider two identical households each with two adults and two children and total household consumption of $120. Assume one household lived in a country that used adult-equivalence adjustments that assume a child consumes half that of an adult. In this country, the individuals of this household would be assigned $40 as their measure of consumption (ie. $120 divided by 3 adult equivalents). Assume the other household lived in a country that defined individual level wellbeing based on a per capita measure of consumption. This household would be assigned $30 as their measure of wellbeing. While both households are identical, the household living in the country that used an adult equivalence adjustment for allocating household measures to the individual would appear to be significantly richer. For a description of adult-equivalence adjustments, see James and Schofield (1990). 17 The adoption of a per capita scale imposes cross-country comparability and is easy to explain. It does not, however, address the deeper issue of what the "best" equivalence scale might be in each country. 18 Zero consumption values may exist in household survey data, but they are often times treated as nonresponders, non-compliant responders and are frequently deleted from the file. 13 employment and mostly agricultural activities, it can be very difficult for survey respondents to provide reasonable answers to income questions. In contrast, in richer countries where work week patterns are uniform, income might be relatively easy to collect while collecting data on consumption may well suffer bias from respondent fatigue. The "best" concept with which to measure household wellbeing may therefore depend on context -on the nature of the income and consumption flows in each society.
Mexico, for which income is used to measure poverty in PovcalNet, provides a useful example of the difference. PovcalNet includes both income and consumption measures for Mexico for the same years. Figure 2 compares the cumulative density functions for both income and consumption, and makes it very clear that poverty estimates differ between the two measures of wellbeing. Using the older 2005 PPP data and $1.25 per day poverty line, the poverty headcount based on income was 3.3 percent, while the estimate based on consumption is 1 percent. It is noteworthy that the consumption distribution dominates income over the entire range -at every candidate poverty line, poverty as measured by income is greater than it would be if measured by consumption. By no means is this finding uniform across countries that collect income and consumption: the relative rankings of the two measures vary substantially across countries.
Indeed, in part because there are few or no fixed, observable patterns in the relationship between income and consumption, no adjustments are made to income or consumption distributions in PovcalNet to make them more comparable. While the issue of zero incomes is not currently a very large percent of the global poverty count, it does have significant implications for poverty projections. Projections are based on increasing over time the measure of wellbeing (either consumption or income) by some estimated growth rate. If income is used as the measure of wellbeing, a growth rate applied to zero income does not change the income level. Projections by construction will never grow the zero incomes out of poverty. This is particularly relevant when examining poverty projections for Latin America, which predominantly uses income data. While poverty declines significantly over time in most regions, the projected poverty rate in Latin America is quite stagnant due in large part to this issue of zero values.

[INSERT FIGURE 2 APPROXIMATELY HERE] [INSERT TABLE 2 APPROXIMATELY HERE]
Issues of comparability are not limited to whether income or consumption is the welfare variable of choice. Research in data collection methods has established that factors such as the length of the recall period and the number of food items listed has a large effect on the resulting measure of estimated food consumption. 19 One prominent example of differences in recall periods having large effects on measured poverty is described by Deaton and Kozel (2005a) The World Bank global poverty estimates for 2012 in contrast continue to use the uniform 30-day recall period in the headline estimates for India, but also report on the difference in the poverty count when using a modified mixed recall period (MMRP). 21 The MMRP contains a shorter, 7-day recall period for a subset of food items, which leads to much higher estimates of consumption and therefore lower poverty estimates. For 2011, the estimates based on the URP suggest a poverty estimate of 21.2%, while the MMRP suggest significantly lower poverty rate at 12.4%, which is a difference of more than 100 million people being moved in our out of the current poverty count depending on the recall period in one country's survey.
In the case of India, the change in the recall period creates a challenge for making comparisons over time. But the challenge is not limited to India alone, nor just to changing questionnaires over time. Smith, Dupriez and Troubat (2014) assess recent questionnaires for national household surveys from 100 countries, and focus on how data on food consumption is collected. They document large variation in all dimensions of survey implementation from mode of collection, duration of fieldwork, to coverage of food items. They report, for example, that 30 percent of the countries have a recall period for food consumed that is greater than two weeks, while 41 percent have recall periods of less than a week. In comparing extreme poverty based on the $1.90 line across countries, one is implicitly assuming that despite the variation in questionnaire design, consumption is measured comparably.
In order to shed light on this assumption, Beegle et al. (2012) carried out a careful experiment in Tanzania to examine how changing several different aspects of the questionnaire affects measured consumption and, thereby, the estimates of poverty. They administered eight different questionnaires, designed to elicit the same information on consumption, to random subsamples from Tanzania. They tested a variety of relatively common ways of collecting information about consumption, contrasting diary with recall, shorter recall periods with longer recall periods, and varying levels of disaggregation of the listed commodity items. One test design included a long list of consumption items and a 7-day recall period. Another test design maintained everything the same, but switched the recall period to 14 days. 22 This simple change resulted in a reduction in mean consumption by about 12 percent, and an increase in poverty by 8 percentage points. Similarly, maintaining the 7-day recall period, but collapsing the food items prompt into fewer, broad food category questions, produced a mean consumption estimate that was 28 percent less than the personal diary module. 23 These findings are not unique to the population examined by Beegle et al. (2012). Jolliffe (2001) presents findings from another between-groups experiment where two subsamples from the population of El Salvador were administered different questionnaires. Both questionnaires were designed to elicit the same information, but one was longer with more prompts for different consumption items. The other, shorter instrument, asked about all the same food types, but collapsed the prompts into broad food categories. 24 The recall period was the same, as well as other aspects of the instrument. The key finding from the experiment is that even though the samples were the same in all relevant dimensions, due to the differences in the questionnaire design, the short-questionnaire sample had a measured poverty rate that was 46 percent greater (due to lower measured consumption) than the long-questionnaire sample.
An important implication of these studies is that differences in household survey questionnaires over time (and between countries) will result in limited comparability in consumption and absolute poverty levels. This finding is true whether the differences is in how the survey is administered (for example, diary or enumerated), the extensiveness of the survey (short but comprehensive list of items compared to long list of items), the recall period (for example, 7-day compared to 14-day), or time of interview (for example, whether in the lean or harvest season). For more discussion on this point, and the importance of data quality in general, see Bamberger et al. (2006), Biemer and Lyberg (2003), and United Nations Statistical Division (2005).
Finally, although all the distributions in PovcalNet originate from household surveys, PovcalNet's estimation do not always use this microdata directly. In cases where microdata are not available or when the World Bank has not been authorized to share these data publicly, PovcalNet uses grouped data (such as quintile or decile shares) and poverty is estimated using a parameterized Lorenz curve, following the computational procedures suggested by Datt (1998), applying methods suggested by Kakwani (1980) and Villasenor and Arnold (1989). 25 Although studies have found that such methods based on grouped data do fairly well at estimating poverty at low poverty lines, consistently and without bias, microdata is generally more accurate and ensures consistency with national data sources (Minoiu and Reddy, 2009). A strength of the current update is therefore that it relies on a much larger share of microdata for recent poverty estimates (after 1990) than previous versions of PovcalNet: grouped data is now used only for six of the 131 countries covered. 26

Inter-temporal price adjustments
Because prices can change rapidly, inter-temporal price deflators are essential for comparisons of real standards of living over time. Since official PPPs are only available for the benchmark years of the ICP, inter-temporal price deflators are necessary to adjust welfare measures for changes in prices between the survey year and the ICP benchmark year. 27 Similarly, to estimate an international poverty line based on national poverty lines, inter-temporal price deflators are used to adjust the value of national lines from various years to the ICP benchmark year, before being converted to PPP-adjusted dollars.
The choice of inter-temporal price deflators can have considerable impact on estimates of global poverty because it affects both the PPP-adjusted values of income and consumption, as well as the value of the international poverty line. Jolliffe and Prydz (2015) provide an example where moving from using only official CPI data on inflation to the inter-temporal price indices used in PovcalNet leads to an increase in the global poverty line of $0.12, which implies an increase in the global poverty count of more than 100 million people. While PovcalNet primarily uses CPI data, it does also use different indices for a few countries, and this choice alone has a substantial impact on the count of poor people. Given that the poverty estimate is sensitive to the choice of price deflators, the decisions made about these indices are important. Table 3 shows that for more than 75 percent of the countries, the World Bank poverty estimates use the annual official national consumer price index (CPI) reported in WDI to deflate current LCU values from the survey year to constant prices of the ICP base year. 28 However, in cases where surveys take place in specific months, or where inflation is high, PovcalNet uses monthly CPIs from official sources instead. These series are reported to PovcalNet by World 26 The countries for which grouped data is used for the 2012 estimate are China, Iran, Maldives, Trinidad and Tobago, and St. Lucia. 27 Alternatively (and equivalently), the international poverty line can be converted to local currency in the ICP base year, and deflated by a temporal deflator to the survey year. The World Development Indicators (WDI) does provide extrapolated PPPs for private consumption using official CPIs, but these are not used in poverty measurement, in part due to the decision to use inter-temporal measures of price change that differ from the WDI CPIs for a selection of countries. 28 The CPI series reported in World Development Indicators originates from the International Financial Statistics data files from the International Monetary Fund (IMF). This series is also used in WDI to extrapolate PPP conversion factors to years other than the ICP base-year.
Bank country and regional teams, citing official national sources. For example, from the LAC region, CPIs are primarily monthly series since survey fieldwork is undertaken during specific months of the year. For China and India, PovcalNet uses separate consumption distributions from rural and urban areas and therefore also use these countries' urban and rural CPIs as deflators. Indonesia also reports separate rural and urban distributions, but no disaggregated deflator is available, so the national CPI is used.
Last, for a small number of countries (seven), PovcalNet uses alternative inter-temporal deflator estimated from price data in household surveys or other alternative sources. These are all countries where there is either no CPI data or there are concerns about the accuracy of the data for measuring changing price levels as experienced by poor people. Using Tanzania as an example, Sandefur (2013) illustrates that there are large differences between CPI data and household-survey-based estimates of price changes. Gimenez and Jolliffe (2014)  Ghana, Lao, Iraq, Malawi and Tajikistan. Annex 1 provides an overview of these countries and the deflators used. For these countries, we have made the assessment that the alternative deflators provided a more accurate picture of changes to prices, particularly to those faced by poor people.
As one example of the importance of these differences, in the case of Tajikistan using the official national CPI would imply an 18% annualized growth of household survey mean from 2004 to 2009, a growth rate that was considered to be unreasonably high. The alternative inter-temporal deflator used for national poverty measurement implies 6% annualized growth. [INSERT

Purchasing Power Parity -Updating measures from 2005 PPPs to 2011 PPPs
Income and consumption measures available from national household surveys, as well as national poverty lines, are denominated in local (national) currency units (LCU). To be able to compare the standard of living between countries, one needs to express the measure of consumption or income in common units. One solution would be to use market currency exchange rates to express all values in a common currency, but these exchange rates fail to accurately reflect relative purchasing power, particularly for poorer countries. One reason for this is that non-traded goods, and especially services, are typically cheaper in poorer countries at market exchange rates (Balassa, 1964;Samuelson, 1964). This implies that using market exchange rates to convert consumption or income underestimates the real standard of living in poor countries.
To address this issue, international poverty comparisons use exchange rates based on PPP conversion factors for private consumption from the ICP. These are essentially exchange rates that attempt to adjust for price differences and to ensure that a dollar has the same purchasing power across countries. PPP factors convert the value of consumption from the LCU into a common currency (ie. USD) in a manner that allows for comparability across countries. PPPs actually enter the international poverty estimation at two stages. First, they are used in estimating an international poverty line based on national poverty lines (converted with PPPs) from some of the poorest countries in the world. Second, to assess poverty in each country, PPPs are used to convert the international poverty line into local currencies or, equivalently, to convert consumption and income distributions from local currencies to PPP dollars, for international comparison.
The PPPs from the 2011 ICP imply some significant revisions to relative price levels  and 2011 rounds, does indeed diverge systematically between rich and poor countries. Figure 3 shows Without an explanation for this divergence in price data, and after careful review with country poverty economists, we established a rule that if exceeds two standard deviations from the mean, the poverty estimates would not be based on the 2011 PPP data, but would remain 33 We exclude non-benchmark countries (i.e. countries for which regression-based PPPs are reported) in this analysis, as we assume actual PPPs more accurately capture the price levels than imputed PPPs. 34 This divergence is easiest to observe by looking at rich countries with around 50,000 GNI per capita. There is a large mass of countries with right around one standard deviation below the mean value for the entire sample. It is also useful to note that one standard deviation below the mean is 1.17, and the U.S. has a equal to 1.15. 35 That is a sample of 140 countries. 36 See the discussion in section 2, drawing from Deaton and Aten (2014)

Within-country spatial price adjustments
At any given time, prices vary not only between countries, but also within countries. To adjustments to the PPPs for India and China were also motivated by a concern that the 2005 ICP price collection was heavily concentrated in urban areas and that the national PPPs therefore possibly had an urban bias (Ravallion and Chen, 2010;Chen and Ravallion, 2008;Ravallion, 2008).
For these three countries, national PPPs from the ICP are disaggregated into rural and urban PPPs that are constructed to reflect cost-of-living differences between rural and urban areas. In backing out separate rural and urban 2011 PPPs for these three countries, we follow steps that are conceptually equivalent to those used by Ravallion (2008, 2010). To unpack the national PPP into an urban and rural PPP, this method uses rural and urban poverty lines to measure differences in the cost of living, as well as information on the rural and urban shares of the ICP price collection. 40 Following notation from Lakner et al. (2015), the adjustments can be summarized as follows. Let be the relative urban-rural difference in cost of living (expressed as the ratio of urban and rural poverty lines, / ) and , the urban share of total price outlets in the ICP price collection for each country. The PPPs are then unpacked into 39 One distinction though is that household survey weights expand to the population of individuals, PPP price data is weighted by the volume of total expenditures in the country. 40 Because cost-of-basic-needs poverty lines are typically estimated to include an allowance for nonfood consumption, and this allowance frequently differs across urban and rural areas, differences in poverty lines are more accurately described as reflecting differences in the cost of living, rather than differences in prices: they potentially reflect differences in both prices and quantities. In contrast, the adjustments for LAC and ECA countries, discussed below, are more accurately referred to as spatial-price adjustments.
urban and rural PPPs such that their ratio is equal to , the urban-rural difference in the cost of living. ( This of course only defines the ratio of the PPPs. To establish unique values for PPPU and PPPR, it is also necessary to know the magnitude of the difference between the urban and rural measures. To this end, the urban share of price outlets, , is used as a weight to measure the difference between the urban and rural PPPs.

(4)
Equations 3 and 4 can be solved to derive unique values for and .  Table 4 summarizes the data used in applying these adjustments and the resulting rural and urban 2011 PPPs for China, India and Indonesia. 42 In addition to cost-of-living differences, this difference also reflects differences in the definition of consumption expenditures in the rural and urban surveys. 43

[INSERT TABLE 4 APPROXIMATELY HERE]
Within-country spatial price adjustments in LAC and ECA The second category of within-country spatial adjustments is pertinent to the Latin America and Caribbean (LAC) region, and the Europe and Central Asia (ECA) region. For most countries in these regions, PovcalNet draws on spatially adjusted consumption and income aggregates from regional databases. In LAC, PovcalNet uses aggregates from the Socioeconomic Database for Latin America and the Caribbean (SEDLAC) in which all rural incomes are increased by a factor of 15% to capture differences in rural-urban prices, based on average ruralurban price differences observed in the region (CEDLAS and World Bank, 2010). Although this uniform adjustment is somewhat arbitrary, CEDLAS argues that it is preferable to ignoring the problem of spatial price differences or allowing for different methods across countries. Adjusting all incomes from SEDLAC to urban price level is particularly reasonable in the context of international poverty measurement using PPPs, since the ICP exercise (both in 2005 and 2011) in this region is carried out solely in urban areas and therefore explicitly does not capture rural price levels. 45 Similar to the LAC region, PovcalNet also draws its consumption and income aggregates in ECA from a regional database, ECAPOV. This regional database, maintained by the World Bank, also embeds spatial-price adjustments into many of its consumption aggregates. ECAPOV adopts, whenever information on households' purchased quantities is available, a spatial Paasche index based on unit values from food consumption, which is then applied to all consumption items (including non-food items). Unlike the LAC region, however these prices are not expressed at the urban level, but adjusted to reflect national prices. 46 Table 5 recaps the various spatial price adjustments embodied in PovcalNet data. From the perspective of comparability, it would be desirable to have more consistent methods to adjust for within-country differences in prices. As one example, the adjustments for rural-urban differences in the cost of living made for China, India and Indonesia differ from those made for 45 Maintaining these adjustments is also helpful for consistency between regional and international measurements of poverty. 46 Another way of stating this is that in ECAPOV, mean consumption levels are not changed when adjusting for spatial differences in prices.
other countries, and the overall poverty count is sensitive to the assumptions underlying these adjustments. Jolliffe and Prydz (2015) note that the approach followed for these three countries has the overall effect of lowering the poverty estimates, relative to a benchmark of simply applying the national PPPs to national distributions.
While further research is needed to help identify a more uniform approach to accounting for within-country, spatial-price differences when estimating global poverty, we believe the approach followed for the 2012 estimates offers two desirable properties. First, this approach allows us to separately estimate the extent of rural poverty in each of these countries, which is useful given the sheer size of the rural population in each of these countries. Second, and perhaps most importantly, the approach we follow for adjusting the 2011 PPP factors for China, India and Indonesia is the same approach followed by the World Bank for the previous estimates based on the 2005 PPP conversion factors. The only change made is that the rural-urban scaling factors have been updated to reflect more recent data. This approach allows us to continue monitoring progress towards the goal of eliminating extreme poverty using the same yardstick as previously used. For countries for which there is no survey available in the specific reference year for the global estimate, the country-level poverty count is estimated by extrapolating consumption or 47 Kraay (2006) shows that growth is one of the key sources of poverty reduction.

28
income from the latest survey using growth rates from national accounts. This procedure assumes distribution-neutral growth, i.e. no change in inequality. In addition, because growth in survey means has historically been lower than growth observed in national accounts data, the growth rates used for extrapolating the survey data are adjusted for these observed differences. 48 If a survey is available after the reference year, that information is also used: An average of the two extrapolated estimates is then reported, weighted by their relative distance from the reference year. The mechanics of the extrapolation and interpolation are described in more detail in Ravallion (2003), Chen and Ravallion (2004) and Jolliffe et al. (2014).
For countries without any poverty estimates, extrapolation to the reference year is of course not an option, and PovcalNet does not report poverty estimates for these countries.
Nonetheless, these countries do indirectly enter the regional and global poverty counts. The country-level poverty headcount used to account for them in regional totals is the product of the estimated poverty headcount in the remainder of the region to which they belong (e.g. sub-Saharan Africa) and the country's own population, as reported by the WDI. So, when aggregating to regional and global totals, developing countries without any poverty measures are in effect assigned the average headcount for their regions. This procedure is obviously not intended to suggest any inference about poverty in those specific countries, which is why numbers are not reported for them. Rather, it is a compromise intended to avoid undercounting poverty in those regions with weaker data coverage. 49 High-income countries are assumed to have no people living in extreme poverty in our global count. Although some people in rich countries report household per capita incomes that are below the international poverty line, per capita consumption is above this threshold for nearly everyone. Drawing an example from the United States, Chandy and Smith (2014) find that 1 to 4 48 For most countries, extrapolation is done using the household final consumption expenditure (HFCE, also known as private consumption expenditures, PCE) component of national accounts. For the Sub-Saharan Africa region, overall real GDP growth rate is used due to poor or inconsistent measurement of HFCE component. Historically, across developing countries, growth in household survey means has averaged approximately 87% of the consumption growth from national accounts (Ravallion, 2003). Therefore, when applying national accounts growth rates to modify survey means, an adjustment factor has been used: 0.87 for all countries, with two exceptions. Those exceptions are China and India, for which additional specific research has been conducted. For China, we use 0.72, and for India 0.51. Note that, for China and India, survey data is available in the majority of the reference years, including in 2012, and no such extrapolation takes place for these two countries. 49 For an overview of countries missing poverty data see Serajuddin, et al. (2015). percent of the population live below $2 a day when using income measures, but that less than 0.1 percent live below this threshold when using consumption measure. While the assumption of zero people in extreme consumption poverty may not be fully supported, it is a useful simplifying assumption that appears to closely approximate the correct estimate.
In most cases, our estimate aims to be as complete as it is treated as providing coverage for the estimate. If the coverage rate is less than 60% we do not report the regional poverty estimate. For reference year 2008 and later, the Middle East and North Africa region falls below this threshold and poverty numbers are not reported.

Updating the international poverty line based on 2011 PPPs
In April 2013, the World Bank Group announced the goal of ending extreme poverty, which was defined specifically as reducing "the percentage of people living with less than $1.25 a day […] to no more than 3 percent globally by 2030. Another desirable attribute is that it is a relatively simple adjustment to make. This decision ensures that the concept (poverty lines of the same 15 poor countries); method (simple average across these same 15 countries); and data (the same 15 national poverty lines) are unchanged. The only element that is changing is the PPPs used to convert the national lines to estimate an international line. However, using the average of these 15 countries does have potential limitations, in part due to weak statistical support, as reviewed by Jolliffe and Prydz (2015). Deaton (2010)   Not unexpectedly, Table 7 shows that given the regional differences in revisions of relative price levels from the 2011 ICP, the average "equivalent poverty lines" diverge regionally. Sub-Saharan Africa's "equivalent poverty lines" give a simple average of $ would lead to small changes in countries in Africa, higher poverty in countries in LAC and ECA, while for countries in EAP and SAR poverty would fall with the new PPPs. Of course, the precise changes to the regional aggregates will depend on population shares and the density of the distribution around the poverty lines in these countries, as well as other adjustments.
[ All in all, most of the alternative approaches that have been proposed for updating the international poverty line to 2011 PPPs end up generating lines that are either exactly or very close to $1.90 a day. We take this as providing reassurance that our approach -deliberately motivated by a concern with preserving the integrity of the goalposts used for the SDGs and the Bank's own twin goals -is in fact reasonably robust.

Results
This PPPs and the $1.90 line in our analysis. In addition to the findings in this paper, more extensive analysis of recent trends in global poverty, and the outlook for the future, is available in World Bank (2015b) and Cruz et al. (2015).

Updated country-level estimates of poverty
Although

The regional poverty profile with the 2011 PPPs
Globally, we estimate that 897 million people, or 12.7 percent of the world population, live in extreme poverty in 2012 under the $1.90 poverty line at 2011 PPPs. The highest regional poverty rate is in Sub-Saharan Africa, where 42.7 percent of the population is estimated to live below the extreme poverty line, followed by South Asia (18.8 percent) and East Asia (7.2 percent). Table 8 summarizes the global and regional poverty measures (percent poor, depth of poverty and number of poor) for a selection of reference years between 1990 and 2012. 59 In contrast to some of the large country-level revisions, the regional poverty profile based on the change to overall poverty rates in these regions appears small at the $1.90 line, the changes are more evident at higher poverty lines. As discussed, because of limited coverage, we do not report regional numbers for the MENA region at this time, although they are used in calculating the global average. 61

Past progress and future prospects for ending poverty
Estimates for selected reference years using both the $1.25 line and $1.90 lines from 1990 until 2011 paint a very similar picture regarding the evolution of global and regional poverty. Global poverty has declined rapidly from 1990 until 2012, driven by a particularly strong decline in East and South Asia. Under both the previous and the updated poverty estimates, the first Millennium Development Goal of halving the share of the world population 60 For India, new PPPs and the revised spatial adjustments work in opposite directions. Revised PPPs suggest that as a whole, India is richer than we used to think. Adjusting only for the new PPP, lowers the poverty rate nationally. Revised spatial adjustments (that are based on India's official poverty lines) suggest that the gap between urban and rural areas in the cost of living is half of what was implied in the earlier official poverty lines. Effectively, rural areas are not so much cheaper to live in than urban areas, as we used to think. 61 Although the regional MENA numbers and several country-level poverty estimates are not reported due to low coverage and concerns with the aggregates, poverty estimates for these countries (and the region as a whole) are calculated for the purposes of global poverty estimation. 38 that live in extreme poverty between 1990 and 2015 was reached well ahead of time. In fact, our most recent results for 2012 estimate of 12.7 percent global poor corresponds to a reduction in the global poverty rate by nearly two-thirds compared to 1990. In absolute terms, the global number of poor has fallen by more than half, from more than 1.96 billion people in 1990 to 897 million in 2012. Figure 8, which shows the evolution of the regional poverty rates under both the  Figure 9, showing that the vast majority of poor people will live in Sub-Saharan Africa (more than 85 percent of the total number) with projections using both lines. Broadly speaking, this analysis suggest that the goal of ending extreme poverty by 2030 will be very ambitious, especially in Sub-Saharan Africa.

Conclusions
The measurement of extreme income poverty for the world as a whole has attracted considerable interest over the last two and half decades. The exercise of updating the international poverty line when PPPs change -and the associated process of reviewing temporal and spatial price adjustments in over one hundred countries -are complex. Transparency and replicability are essential to ensure credibility, and 41 we believe that this paper is a step in that direction. Even more importantly, the PovcalNet online tool will soon include the updated distributions, poverty line and PPP options to enable other researchers to replicate our results, as well as making different choices of their own.
Looking forward, we feel that one area where additional research is particularly needed, in order to enhance international comparability, is that of spatial cost-of-living adjustments.
These are still made in a piecemeal and uncoordinated fashion across regions. While we feel that each of the adjustments we have made here is defensible in its own right, their consistency across space and time requires additional scrutiny. In that matter, and in many others, we look forward to advice from the Global Commission on Poverty, led by Professor Sir Anthony Atkinson, which was recently convened to advise the World Bank on how to take global poverty measurement forward, after this 2012 update.            the annualized growth in household survey mean is 6 percent. Using the official CPI the annualized growth in household survey mean would be 18.5 percent, which is viewed by those with knowledge of the country as unreasonable, in part based on the evidence that GDP growth was assessed at 2.6 percent over that period. The GDP deflator for final household consumption expenditure available in WDI reflects a rate of inflation similar to the deflator used for poverty measures by PovcalNet. See World Bank (2009, Annex 4) for more details.

Tables
The full set of CPIs used is available in the online version of PovcalNet.