Poverty-Specific Purchasing Power Parities in Africa

The paper revisits the issue of poverty-specific purchasing power parities (PPPs), using the most recent (2011) International Comparison Program (ICP) results. The World Bank's global poverty count uses a common international poverty line -- currently $1.90 at 2011 international prices-based on the ICP PPPs for consumption. The use of these PPPs is often criticized for two reasons. First, the ICP PPPs are based on patterns of aggregate household consumption, not the consumption of the poor. Second, the basket of goods and services used for collecting prices for the ICP is not poverty specific. On the first issue, using data from 28 African countries, the paper concludes that the poverty-specific PPPs estimated with household expenditure survey weights are very similar to the ICP PPPs. On the second issue, poverty-specific PPPs were estimated after removing items deemed to be irrelevant for the poor. The overall effect of removing these items from consumption PPPs is shown to be negligible.

The paper revisits the issue of poverty-specific purchasing power parities (PPPs), using the most recent (2011) International Comparison Program (ICP) results. The World Bank's global poverty count uses a common international poverty line-currently $1.90 at 2011 international prices-based on the ICP PPPs for consumption. The use of these PPPs is often criticized for two reasons. First, the ICP PPPs are based on patterns of aggregate household consumption, not the consumption of the poor.
Second, the basket of goods and services used for collecting prices for the ICP is not poverty specific. On the first issue, using data from 28 African countries, the paper concludes that the poverty-specific PPPs estimated with household expenditure survey weights are very similar to the ICP PPPs. On the second issue, poverty-specific PPPs were estimated after removing items deemed to be irrelevant for the poor. The overall effect of removing these items from consumption PPPs is shown to be negligible.

ICP and PPPs
The ICP is a worldwide statistical initiative to collect comparative price data and compile detailed expenditure data of the world's economies. The program's main outputs are PPPs of countries' gross domestic product (GDP) and its main expenditure components. 2 ICP price data are collected via specially designed price surveys, while expenditure values in local currency are compiled from countries' national accounts.
The latest ICP comparison to date is the 2011 ICP, which took place six years after the 2005 ICP. In this sense, ICP comparisons have historically occurred at infrequent time intervals, though this will change starting with the forthcoming 2017 ICP and onwards. 3 ICP comparisons are made from the expenditure side of the national accounts. PPPs are therefore calculated for different expenditure levels of aggregation, starting with basic headings and up to GDP. To maintain consistency with expenditures on GDP, ICP items underlying PPPs at each expenditure level were selected with the idea of approximating the full range of goods and services making up each expenditure level.
The different expenditure levels for which PPPs are calculated are illustrated by Figure 1. The second row from the bottom refers to the basic heading level, which is the building block for the ICP exercise. It 1 Before 2001 the World Bank set its two previous IPLs using ICP PPPs for gross domestic product.
2 Here and throughout the rest of the paper we use the term "economy" interchangeably with "country" to refer to territories for which authorities report separate statistics.  Against this background, it is worth expanding on the topic of how PPPs are estimated. The process, at least broadly speaking, involves two stages. At the first stage, price relatives for individual items are averaged to obtain basic heading level PPPs using the Country Product Dummy (CPD), first formulated by Summers (1973). 4 The Country Product Dummy (CPD) method is carried out within each basic heading by regressing the logarithm of observed country item prices on item and country dummies. In actual computations, the CPD formulation with weights (CPD-W) is used. 5 However, the version of the CPD-W used in the ICP incorporates information on the relative importance of items in a country's consumption rather than actual weights: weights of 3 and 1 are assigned to important and less important items, respectively. 6 At the second stage, basic heading PPPs are aggregated using the GEKS-Fisher method to produce above-basic heading PPPs. The method uses the Fisher ideal index to construct bilateral PPPs for each pair of countries, using basic heading expenditure weights from each country in turn. The bilateral PPPs are then averaged using the Gini-Éltető-Köves-Szulc (GEKS) approach to arrive at a final vector of abovebasic headings PPPs, containing one PPP for each country relative to the numeraire country. For more details on the PPP estimation process and other ICP concepts and methods, see World Bank (2015).

Poverty and PPPs
Because of the need to adjust for price differences between countries, ICP PPPs have largely been of interest to researchers working on global poverty, including those at the World Bank. The World Bank's interest in ICP PPPs gained notoriety in 1990 when it used the 1985 ICP PPPs for GDP to set its dollara-day IPL. Ever since, the use of ICP PPPs in revisions to the World Bank IPL has generated significant 4 For a more detailed description of the method see, for example, World Bank (2015), Chapter 23.
5 The Eurostat-OECD and CIS regions used the Jevons Gini-Éltető-Köves-Szulc* (Jevons-GEKS*) method rather than the CPD to estimate basic heading PPPs. 6 The decision on whether an item is important or less important is taken at the country level once the ICP price collection is complete.
The procedure is not entirely precise so some subjective judgement is involved. 4 commentary, which in turn has prompted research on the role of PPPs in methods to measure global poverty. Deaton and Dupriez (2011) were one of the first to study the effect of estimating global poverty using poverty-specific PPPs. Using 2005 ICP data, they found that consumption PPPs reweighted to a poverty basis are quite similar to the regular 2005 ICP consumption PPPs that use weights from national accounts. They note that weighting differences are probably not of great importance for estimating global poverty counts. To reach this conclusion, they used household expenditure surveys from 62 poor countries around the world to reweight the 2005 ICP consumption PPPs and produce a set of poverty-weighted PPPs. These PPPs were then used to calculate new IPLs and global poverty counts according to various definitions. Similar research at the regional level was conducted by the Asian Development Bank (ADB). The Asian Development Bank (2008) used price data for 16 Asian countries to compile a set of povertyspecific PPPs. To that end, a separate price collection, using modified items from the 2005 ICP, was organized in 2006. The study examined whether data collected on prices for the items that were considered typical of the consumption patterns of the poor would produce significantly different poverty PPPs. This research concluded that indeed PPPs could change more substantially when using items consumed by the poor and poverty-specific weights; however, it is difficult to compare its results directly to the 2005 ICP because of differences in methodology, timing and geographical scope (number of countries). In particular, because of inclusion of more items with loose specifications, comparability of items across countries could be more problematic in the ADB's study than in the 2005 ICP.

ICP and CPI data
Throughout, we make extensive use of 2011 ICP data, classifications and concepts. Our interest though is on PPPs for consumption, so we only work with basic headings belonging to the consumption component of GDP.
In section 4 we focus on the reweighted consumption PPPs. The reweighting process was done at the basic heading level and we worked with 108 of the 110 basic headings for household consumption. Our calculations exclude two basic headings that are outside the scope of most household expenditure surveys, namely those corresponding to expenditures in the domestic market by non-resident households and expenditures of resident households when traveling abroad.
In addition to ICP data, computations in section 4 also required data on consumer price indexes (CPIs) of individual countries. 7 CPIs for general consumption were used to deflate the local currency value of the IPL to prices prevalent in the year each household expenditure survey was conducted. This was necessary when calculating poverty-specific PPPs that required identifying households below or around the IPL. All CPI series used were sourced from ICP regional implementing agencies, the IMF Statistics Department, and, to a lesser extent, the World Bank World Development Indicators (WDI).
In Section 5 we turn to the item level. Computations in this section relied on item-level price data and related metadata from the 2011 ICP Africa exercise, which included 50 countries. Unlike in section 4, we used information from all 110 basic headings for household consumption when possible, even if the two basic headings excluded in section 4 contain no item-level information.

Standardized household expenditure survey data
Our reweighting of consumption PPPs required switching regular ICP national accounts sourced expenditures with those from household expenditure surveys. In particular, we used expenditures from a set of standardized data sets derived from existing household expenditure survey data files. Table 1 lists the survey title and year of the household expenditure surveys underlying each of the 28 standardized data 7 CPIs measure the average change over time in the prices of consumer goods and services purchased for consumption by a reference population.
POVERTY-SPECIFIC PURCHASING POWER PARITIES (PPPs) IN AFRICA 5 sets used. In addition, it provides information on the number of survey items and ICP basic headings in each standardized file.
While we aimed to cover and reweight PPPs for all countries in Sub-Saharan Africa, we were limited to the 28 countries with available standardized data sets at the time of our study. 8 We must underline that the feasibility of much of our results in section 4 is reliant on these standardized data sets. In this sense, it is useful to have some understanding of the standardization process and its output.
The standardization process provides household expenditure survey data in a more accessible and manageable format-a considerable feat considering the lack of harmonization of household expenditure surveys across countries. The process utilizes a common data dictionary (i.e., common variable names, formats, and data structures) to extract consumption expenditure data from household expenditure surveys. The process involves three main steps that we summarize as follows: (i) mapping survey items to ICP basic headings, (ii) annualizing consumption values; and (iii) identifying and fixing outliers.
The first of these steps involves mapping survey items from each household expenditure survey to one of the 110 ICP 2011 consumption basic headings. This ensures some correspondence between the survey data and the ICP. 9 The second step is required because household expenditure surveys collect data using 6 recall periods that vary depending on the type of goods and services. Finally, the third step focuses on detecting and fixing top outliers in consumption values to further ensure the reliability of the survey data. For more details on the standardization process, see Dupriez (2007).

Method and approach
The usage of national accounts instead of poverty relevant weights in calculating consumption PPPs is a typical critique of the IPL used for global poverty measurements. However, as mentioned earlier, this was not a problem as found by Deaton andDupriez (2011) using 2005 ICP data. In revisiting this topic, we follow up on Deaton and Dupriez (2011) and calculate poverty-reweighted PPPs using 2011 ICP data. The reweighting procedure involves substituting the conventional national accounts-based basic heading expenditure weights with poverty-relevant weights from household expenditure surveys. Implicit in this process is the requirement to identify poor households consistently across countries, which in turn leads to a well-recognized circularity issue: PPPs determine the IPL below which the poor live, whose expenditure weights in turn affect the PPPs. We solve this issue using an iterative procedure to arrive at a final set of poverty-weighted PPPs. The procedure can be described as follows: the nth iteration would involve estimating a new set of PPPs-PPP n -with some previous set of poverty weights w n-1 and then use PPP n to estimate a new set of poverty weights w n to be used in the next iteration n+1.
In this context, we estimated a variety of poverty-weighted PPPs using different methods to identify poor housing across household expenditures surveys. In particular, we used the uniform and bi-weight kernels to obtain the consumption patterns from households around the IPL. We then derived average basic heading level expenditure weights using a democratic method. 10 As a control, we also ran the computation for poverty-weighted PPPs using expenditures from households below the IPL and averaged the basic heading weights using a plutocratic method.
In discussing the poverty-weighted PPPs based on iterative procedures, the issue of uniqueness of the solution deserves special attention. Deaton and Dupriez (2011) note that there is no guarantee that a unique solution exists in the general case. They report that uniqueness is guaranteed though for log-linear budget shares and the Törnqvist index.
However, it seems that uniqueness is also guaranteed in the general monotonic case for budget shares, not only log-linear, and even for a non-monotonic relationship without sharp oscillations. In practice, given actual data, there seems to be no problem with convergence to a unique solution, or, at least, cases of multiple solutions have not been discovered.

Convergence speed of iterations and kernels
In general, the convergence to a unique solution was found to be extremely fast, but it did depend on the type of filter (i.e., kernel) and the bandwidth (bw) employed. On the latter, we explore three different bandwidths for each of the two kernel shapes. 11 Understandably, the uniform kernel is less stable as it is affected by weight irregularities and distribution lumpiness at both ends of the band, whereas for the bi-weight kernel the discrepancies at the ends have almost no effect on the result. As the uniform kernel is actually a band with no re-weighting, 10 Plutocratic weights are the kind obtained from the national accounts, whereby all households are treated as one unit. Weights using this method are derived from the total expenditures of all households on a given basic heading, so richer households exert a greater influence on the computation. Democratic weights represent all households equally and are derived by taking the average of the expenditure shares of each household on a given basic heading, so all households exert an equal influence on the computation. 11 Kernel bandwidths used are from widest to most narrow: 1.0, 0.5, and 0.25. A wider bandwidth increases the width of the shape around the IPL so that more households are included in the sample when extracting expenditures from a household expenditure survey. 7 using this kernel as a comparator would show the effect of re-weighting in a kernel. The convergence speed for all poverty-weighted PPPs are presented in Figures 2 and 3.

FIGURE 2. ITERATION CONVERGENCE OF POVERTY-WEIGHTED PPPS, 28 COUNTRIES
In the case of 28 countries, the fastest convergence, as expected, is produced by the bi-weight kernel based PPPs with bandwidths of 1.0 and 0.5. The same kernel with a bandwidth of 0.25 converges significantly slower. Poverty-weighted PPPs based on plutocratic weights below the IPL ("below IPL PPPs") converge relatively fast in the beginning but then start oscillating. All indexes based on uniform kernels exhibit oscillations as well. It could be noticed that the only country that oscillated with the below IPL PPPs was Guinea. Removal of the Guinea data from the data set produces somewhat more consistent results for all indexes, with the below IPL PPPs discontinuing the oscillations altogether. The results without Guinea price data are presented in Figure 2. In general, the convergence picture is quite similar to the 28 country case. As before, the fastest convergence is produced by the bi-weight POVERTY-SPECIFIC PURCHASING POWER PARITIES (PPPs) IN AFRICA 8 kernels with bandwidths of 1.0 and 0.5. The below IPL PPPs converge relatively fast as well. In this case, the uniform kernel with bandwidth 1.0 converges fully. The only two indexes that produce oscillations are based on the uniform kernels with bandwidths 0.5 and 0.25. The reason why one country could have such an effect lies in the lumpiness (i.e., discontinuity) of actual household survey expenditure data. For example, if a certain number of households in some country around the poverty line exhibited a rather unusual weight structure combined with an uneven density of the probability density function, it would result in oscillations for the below-IPL index. This has nothing to do with the methodology employed, but rather reflects imperfections of the original household expenditure data. Smoothing the input data would solve the problem and at the same time would make the household expenditure data more realistic, without affecting much the resulting PPPs.
How critical are the oscillations in these PPP calculations? It turns out they are quite insignificant. From Figures 2 and 3 we can see that in the worst case scenario-uniform kernel with bandwidth 0.25the oscillations are around 0.01% (!) on average, which is several orders of magnitude better than expected precision of the PPP computation.
It also turns out that all bandwidth of kernel based PPPs produce very similar results, with most countries having results with a standard deviation (SD) of around 0.1-0.2%. 12 Only the Gabon poverty weighted-PPPs have an SD of 0.4% (see Figure 4), with the bi-weight poverty-weighted PPPs systematically higher than their uniform kernel based counterparts (see Table A1 (annex)). This is probably related to some peculiarities of Gabon's probability distribution function.

Effect of removing one country
We have seen the effect of removing Guinea on the convergence of poverty-weighted PPPs, now let us look at this effect on the regular consumption PPPs based on expenditures from the national accounts ("SNA-based PPPs"). The results for SNA-based PPPs are shown in Figure 5. The scale is intentionally 12 SD is estimated with respect to the regional unweighted geometric mean in order to remove the base country effect. kept the same as in Figure 4 for easy comparisons. The average effect of removing one country on the rest of countries is around 0.11% in this case. Next, we examine the effect of removing Guinea on the below IPL and kernel based poverty-weighted PPPs. Figure 6 shows that the below IPL and kernel based poverty-weighted PPPs change insignificantly in this case, with the below IPL PPPs being more stable. All kernel based PPPs display very similar magnitudes of the effect, with the biggest effect being for Rwanda (for the 27 country case, the SD of the effect was 0.19-0.23%). Interestingly, Rwanda's PPP was quite stable under the below-IPL PPP computation.

Effect of moving from below IPL PPPs to kernel based PPPs
Once we established that all kernel based PPPs were quite similar, and the country-removal effect was of a quite limited importance, the next step would be to study the effect of moving from the below IPL PPPs to the kernel based PPPs. This effect is presented in Figure 7, with the biggest outlier being Rwanda with a 3.2-3.8% difference. We can also see that the SD of the differences in country poverty-weighted PPPs is 1.0-1.1% depending on the kernel. Those numbers again exceed the expected error in the ICP PPP computation.

Effect of poverty-weighted PPPs on poverty rates
With the poverty-weighted PPPs according to various definitions being that close to one another, the country poverty rates they generate are quite close as well. These poverty rates are presented in Table A3 (annex). We can see that all kernel based poverty weighted PPPs produce virtually identical poverty rates. Those poverty rates are also quite close to those originating from the below-IPL PPPs.

Effect of moving from SNA-based PPPs to poverty-weighted PPPs
Finally, we are going to look into the effect of going from SNA-based PPPs to poverty-weighted PPPs. This effect is presented in Figure 8. One thing to note is that the effects in Figure 8 are significantly larger than those presented earlier. The resulting SDs for the various types of poverty-weighted PPPs estimated are 2.7-3.1%. These numbers are still significantly better than the expected precision of ICP PPPs (a 5-10% range). Again, all the kernel based PPPs and the below IPL PPPs are quite close to each other.

Method and approach
We now turn our focus away from the basic heading level and toward the item level. The povertyspecific PPPs in the previous section attenuated the influence of basic headings irrelevant to the poor through reweighting. However, if we want to produce PPPs untainted by any possible influence of items irrelevant to the poor, then the reweighting process is by and large effective, but by no means sufficient. Items seldom consumed by the poor, such as extra virgin olive oil, would still be present within basic headings like "other edible oils and fats"; which, overall, have an important role in the consumption basket of both poor and non-poor.
As mentioned before, the inclusion of items that are arguably irrelevant to the poor in ICP consumption PPPs has often raised doubts on their applicability for poverty analysis. However, the effect (if any) of constructing consumption PPPs that exclude the price of, say, extra virgin olive oil or Kellogg's Cornflakes is not immediately evident.
In light of this, we produce new consumption PPPs with items deemed irrelevant to the poor removed from the calculation. We refer to these poverty-specific PPPs as reduced-list PPPs and produce three scenarios: (1) after removing items priced only in supermarkets, (2) after removing clothing and footwear items belonging to a medium or high brand stratum (for brevity, we name them "branded garments & footwear"); and (3) after removing food and nonalcoholic beverage items that we categorized as premium beforehand.
Within each reduced-list scenario, we computed two sets of PPPs for each of the 50 African countries. First, a full-list set, based on the full basket of items used for collecting prices for the 2011 ICP in Africa, and, second, a counterfactual reduced-list set after removing items deemed irrelevant according to each scenario. 13 Basic heading PPPs for the two sets, in each of the three scenarios, were estimated using the 13 Full-list PPPs were only estimated for expenditure categories for which counterfactual PPPs were also produced. Otherwise, published 2011 ICP PPPs were used. Hence, the full-list PPPs for the 'supermarket only' and 'premium food and nonalcoholic' items are equivalent, but differ slightly from the full-list PPPs for the 'excluding branded clothing & footwear' scenario. CPD method in its unweighted form. For comparison purposes all basic heading PPPs were then aggregated to the level of household consumption using the GEKS-Fisher procedure.
Depending on the reduced-list scenario, the removal of items was contained to either the "food and non-alcoholic beverages" or the "clothing and footwear" ICP expenditure categories. These two categories were chosen because of their importance as basic necessities as well as for practical reasons. It turns out that the product dataset for the 2011 ICP Africa comparison contains quite detailed and harmonized metadata for items belonging to both categories.
Determining what items to remove in each reduced-list scenario was not without its problems. At first, it would seem that the main difficulty lies in the circularity of the task at hand: it is necessary to consistently identify the poor in each country before assessing what items they consume. Yet, while the circularity is indeed an obstacle, it can be solved using the iterative procedure employed in Section 4. Instead, the lack of item-level detail in most of the household expenditure surveys proved to be the more intractable problem is. On this front, we mined the micro data from household expenditure surveys in 28 Sub-Saharan Africa countries and concluded that their item-level detail is generally insufficient to properly distinguish item-varieties and establish a one-to-one mapping with ICP items.
Given this constraint, we chose the practical alternative of using information from the ICP product data set for Africa mentioned above to identify items that one would expect to be outside the consumption basket of the poor. To identify items priced only in supermarkets we used information from the "required outlet type" field provided for each item. In the case of garments and footwear, we stratified items within the "garments" and "shoes and other footwear" basic headings as branded or unbranded by exploiting the available item-level brand stratum information (high, medium, or low). As stated earlier, we grouped and labeled all high and medium garments and footwear as branded for brevity even if some low items also include some brand specification. Lastly, for the premium food and nonalcoholic beverages scenario we removed items with a relatively high (per-unit) price across all countries within each basic heading, and assume a linear relationship between quantity and prices for each item.
We must underline that all three approaches are not without their drawbacks, but we cannot do any better given the current set of data. Nevertheless, market consumer reports by Nielsen (2014) and information from the Food and Agriculture Organization (2015) indicate that the poor in many African countries rarely shop in supermarkets. Likewise, it is not unreasonable to assume that high brand stratum garments and clothing are outside the scope of what the poor consume, at least on average.

Poverty-specific PPPs by reduced-list scenario
To measure the effect of moving from PPPs based on a full-list to PPPs based on a reduced-list, we compute the SD across (normalized) relative differences in country PPPs due to the shift. 14 These relative difference in PPPs are captured by the ratio between a country's reduced-list PPP and its full-list PPP ("PPP ratio"). 15 The SDs of the normalized PPP ratios in each scenario are presented in Table 2, while country PPP ratios by reduced-list scenario are available in Table A5 (annex). 14 All PPP ratios presented are based on normalized country PPPs with respect to the regional geometric mean. This normalization procedure removes the base-country effect that would occur otherwise. 15 A PPP ratio greater (less) than 1.00 denotes an increase (decrease) in country PPP, relative to the region, due to the shift from a fulllist to a reduced-list. Table 2 reveals that the reduced-list PPPs in each scenario are not much different from their full-list analogues, as evidenced by the SDs. The effect of moving from a full-list set of consumption PPPs to each reduced-list scenario is displayed in Figure 9. These results imply that the impact of removing items from the consumption PPPs is negligible for the three scenarios studied. In fact, the SDs of the PPP ratios for each scenario are below the ± 5-10% precision band accepted as target for the ICP PPPs. With this in mind, we proceed to examine each reduced-list scenario in more detail. Figure 9 illustrates the (negligible) effect of moving from a full-list set of consumption PPPs to one excluding supermarket only items. In total, 41 items out of the 367 food and nonalcoholic beverage items were removed.

Effect of moving from full-list PPPs to PPPs excluding supermarket only items
The SD of the relative differences in country PPPs due to removing supermarket only items was 0.88%, as indicated earlier. Swaziland and Senegal had the most changes in their consumption PPPs, but these were still only [+]1.72% and [-]1.62%, respectively.
Regression results in Table A4 (annex) indicate a positive statistically significant relationship between country income and the resulting change in country's PPP after excluding supermarket items. However, the size of the effect is extremely small-it amounts to a 1% price increase over a 12-fold difference in country income levels. The effect indicates that non-supermarket prices in richer African countries may be relatively higher than those in poorer African countries. A possible explanation is that non-supermarket items are specified looser, and poorer countries may be pricing lower quality varieties. In any case, given that the magnitude of the effect is trivial, the impact of any association between income and changes in PPP is close to null.

Effect of moving from full-list PPPs to PPPs excluding branded clothing and footwear items
For this scenario, poverty irrelevant items under the "garments" and "shoes and other footwear" basic headings were removed to calculate the reduced list PPPs. The "garments" basic heading contains 65 items among which 23 are low stratum, while the "shoes and other footwear" basic heading contains 20 items among which 7 are low stratum. Only low stratum items were retained in our calculation of povertyspecific PPPs for this scenario.
The change from a full-list to a reduced-list with only low brand stratum garment items resulted in a SD at the garment basic heading level of 6.33% across all countries, with no specific pattern. At the household consumption level, the SD of the PPP changes across all countries was 0.29%.
The extent of the changes in country consumption PPPs for this scenario can be observed in Figure 9. Botswana showed the largest increase in its PPP at 0.66%, whereas Togo had the largest relative decrease at 0.67%. No significant relationship was found between country income and the resulting change in country's PPPs after excluding branded clothing and footwear items (see Table A4, annex).

Effect of moving from full-list PPPs to PPPs excluding premium food & nonalcoholic beverage items
By assuming that the poor are less likely to consume those items that are relatively expensive across all countries (within each basic heading), we removed 61 out of the 367 food and nonalcoholic beverage items priced in Africa. The effect of removing these premium food and nonalcoholic beverage items on the country PPP ratios is shown in Figure 9.
It is remarkable that the effect of moving from a full-list set of consumption PPPs to one without premium food and nonalcoholic beverages was even smaller in this scenario than in the supermarket only scenario, despite removing 20 more items in the former. As with the other reduced-list scenarios, the effect of this shift was small and practically random in its outcome across countries. This is again evident by the regression results in Table A4 (annex). As a final and general observation for the item-level section, we must add that there could be many factors affecting the price level of the poor and they could even work in opposite directions. For instance, without further analysis, it is impossible to quantify to which degree the poor benefit from economies of scale versus the rich. Similarly, we do not know how item availability in rural areas, where many of the poor live and where many items are not even available, affect the effective price level faced by the poor.

Conclusions
With the new $1.90 IPL and new 2011 ICP PPPs, it was important to see if the Deaton and Dupriez (2011) conclusion on poverty-weighted PPPs being close to ICP PPPs still holds, especially given the changes in ICP methodology that occurred since 2005. This paper found that the conclusion does indeed hold: the deviation between the two is around 2.7-3.1% on average, which is below the expected precision of ICP consumption PPPs of ±5-10%.
In addition, the uniform kernel was employed alongside the bi-weight kernel, to study the effects of kernel shape. Those were contrasted with the below IPL plutocratic consumption PPPs. It was found that all the kernels with various bandwidths produced virtually identical results, and those results were very similar to the PPPs obtained with the below IPL plutocratic index.
All indexes based on kernels and the below IPL PPPs, converged fast in the practical sense, meaning that even though they sometimes oscillated, the degree of oscillation was immaterial. All the indexes employed exhibited a high degree of stability to the selection of countries.
At the same time, the overall effect of removing items from consumption PPPs has been shown to be negligible. Yet, it is important to acknowledge that by using the same set of prices in each of the calculations, it is implicitly assumed that the poor face the same prices as the non-poor. In addition, some critics have pointed out that poverty-specific PPPs should be constructed on the basis of prices paid by the poor. We do not address this issue, since unfortunately studying it fully would require a separate price 15 collection, parallel to the ICP. Instead, we attempted to examine feasible aspects related to the construction of poverty-specific PPPs. Future work at the item-level will explore whether the results from this section can be generalized to other item groupings or groups of countries. Notes: All ratios are based on country PPPs normalized with respect to the regional geometric mean. Source: Authors' calculations.    Notes: Full-and reduced-list PPPs are reported with South Africa as the numeraire country (South African rand=1). PPP ratios are based on normalized country PPPs with respect to the regional geometric mean. For an explanation of normalized PPP ratios see footnote 15.