Abstract
As awareness of the limitations of relying solely on income to measure poverty has become more widespread, attention has been increasingly focused on multi-dimensional approaches, to the point where the European Union (EU) has adopted a multidimensional poverty and social exclusion target for 2020. The rationale advanced is that the computation of a multidimensional poverty index is an effective way of communicating in a political environment, and a necessary tool in order to monitor 27 different national situations. In contrast with the rather ad hoc way in which the EU 2020 poverty target has been framed and rationalized, the adjusted head count ratio applied here has a number of desirable axiomatic properties. It constitutes a significant improvement on union and intersection approaches and allows for the decomposition of multidimensional poverty in terms of dimensions of deprivation and socio-economic attributes. Since understanding poverty as multidimensional does not necessarily require constructing a multidimensional poverty index, on the basis of our analysis we provide a more general consideration of the value of developing a multidimensional index of poverty for the EU.
Introduction
Increasing attention is being focused on multi-dimensional approaches to measuring poverty and social exclusion, identified by Kakwani and Silber (2007) as the most important recent development in poverty research. This is reflected in the fact that the 2020 poverty and social reduction target adopted by the European Union (EU) in 2010 employs a multidimensional measure, combining three indicators (relative income poverty, material deprivation and household joblessness). In justifying this approach the EU Commission (2011) has argued that the computation of a multidimensional poverty index is an effective way of communicating in a political environment, and a necessary tool in order to monitor 27 different national situations.
The academic and policy debates on such methodological approaches have focused on the value of summary indices for communication to a wide audience versus the potentially arbitrary nature of decisions required in combining distinct dimensions. A number of authors have questioned whether acceptance that poverty is multidimensional, necessarily implies a need for a multidimensional poverty index. Ravallion (2011) concludes that it is one thing to recognize that something is missing from a given measure and quite another to conclude that what is required is a single composite index. Nolan and Whelan (2007) note that while a case can be made for a multidimensional approach in seeking to adequately measure, understand and respond to poverty, they are not the same case and one does not simply follow from the other.
Here our aim is to look beyond both the general arguments relating to employing a multidimensional poverty index, and the specifics of the European poverty target and its limitations. 1 Our objective is to consider what can be gained by applying a specific multidimensional approach with clearly understood axiomatic properties, namely the one recently developed by Alkire and Foster (2007, 2011a,b), which allows one to examine in a structured way the implications of key measurement choices for levels of multidimensional poverty, the dimensional profile and the socio-economic processes involved. This approach has been framed primarily in an economic development context, but here we apply it to the countries of the EU making use of newly available and richer comparative data on various aspects of deprivation from European Union Statistics on Income and Living Conditions (EU-SILC).
The adjusted head count ratio approach
Bourguignon and Chakravarty (2003) provide a framework for multidimensional poverty measurement involving both an identification function for counting the number of poor and a poverty measure that combines that information into a statistic summarizing the overall extent of poverty. Axioms analogous to the ones used in the one-dimensional case ensure that multidimensional poverty can be decomposed by dimension and sub-group. The simplest summary measure is the number of dimensions on which an individual or household is deprived, which Atkinson (2003) refers to as the ‘counting’ approach. Atkinson (2003) distinguishes between the union and intersection approaches, the former counting as poor those deprived on any dimension while the latter counts only those deprived on all dimensions. While these approaches are easy to understand, they can be particularly ineffective at separating the poor from the non-poor, with the former tending to identify implausibly large numbers as poor and the latter to capture small minorities.
A key motivation underlying the recent methodological contributions of Alkire and Foster is to address these shortcomings. 2 Their procedure involves a dual cutoff approach. The first relates to the choice of thresholds for individual dimensions. Given a set of deprivation dimensions considered as of equal weight, if a person’s outcome on a given deprivation dimension j exceeds the appropriate threshold zj then the individual is said to be deprived on that dimension. 3 The breadth of each person’s deprivation is simply the number of deprivations experienced. The second cutoff point k is used to determine whether a person is deprived on a sufficient number of dimensions to be considered poor. Thus in order to be multi-dimensionally poor an individual must be above the appropriate deprivation dimension threshold on the requisite number of dimensions.
Following Alkire and Foster (2011a,b), the implementation of the approach is best understood as involving a progression of matrices. The starting point is a set of scores for a group of n individuals on d dimensions. This is the achievement matrix Y. For the second matrix, rather than considering the full range of scores on the deprivation dimensions, we distinguish between those above and below the threshold for each dimension. This produces the deprivation matrix g0 by replacing each entry in Y that is above its deprivation cutoff zj with the deprivation value wj and each entry that is not above the deprivation threshold with 0. Finally, we proceed to take into account deprivation, only for those experiencing sufficient deprivations to be above the second cutoff point relating to number of dimensions. This step involves what Alkire and Foster (2011a) describe as censoring the matrix. Deprivation scores above 0 now relate only to those who are above the specified threshold for the requisite number of dimensions. All others are allocated scores of zero.
The headcount H is the proportion of people who are multi-dimensionally poor. The intensity I is the average deprivation score for those experiencing multidimensional poverty. The adjusted headcount ratio is the product of the headcount by the intensity. Alkire and Foster (2011b) demonstrate that their methodology satisfies a range of desirable axiomatic properties. Of particular relevance for our analysis is decomposability in relation to dimensions and socio-economic groups, that is, the ability to calculate the contribution of each dimension to the adjusted head count ratio and the proportion accounted for by each socio-economic group. The latter estimate of composition may be distinguished from variation in risk levels as reflected in the mean adjusted head count ratio across such groups.
Data and measures
The data employed here come from the 2009 round of EU-SILC, which included a special module on material deprivation. Its availability allows us to explore the dimensionality of deprivation in a more comprehensive manner. Sweden has been excluded from our analysis because of a large number of missing values on the deprivation items, so the analysis covers 28 countries – the remaining 26 EU Member States together with Norway and Iceland. Our analysis of poverty is conducted at the individual level. The total number of individuals included in the analysis is 559,767. However, given that the key deprivation indicators are largely measured at the household level, multilevel analysis of the determinants of such poverty is conducted at the household level employing both household and Household Reference Person (HRP) characteristics. The HRP is the individual responsible for the accommodation. Where more than one person is responsible the oldest individual is chosen. 4 This analysis is performed at the household level with 199,353 cases.
Our analysis makes use of 20 non-monetary indicators of deprivation; where questions have been addressed to individuals we have taken the response of the HRP as applying to the household.
The special deprivation module in EU-SILC 2009 has been analysed by Whelan and Maître (2012), whose factor analysis identified six dimensions of deprivation, namely basic, consumption, household facilities, health of the HRP, neighbourhood environment and access to facilities. The number and type of dimensions identified depends not only on the results of factor analysis, but also on conceptual considerations influencing both the choice of items and the hypothesized relationship between dimensions. Guio et al. (2012) in an analysis that does not include health items identify five dimensions broadly similar to those identified by Whelan and Maître (2012) but involving a somewhat different allocation of the household amenities and basic deprivation items. 5
In what follows we operate with a subset of the dimensions identified by Whelan and Maître (2012). However, our analysis does not depend on establishing the superiority of any particular set of dimensions. It simply requires that the set of dimensions employed are appropriate to the implementation of the adjusted head count ratio approach. Of the six dimensions identified by Whelan and Maître (2012) we exclude the dimension relating to ‘housing facilities’ because a number of the items it includes have close to zero levels of deprivation in the more affluent countries, making comparisons problematic. We also exclude the dimension relating to ‘access to facilities’ because in addition to comprising only two items, it appears to be affected more by urban–rural location than factors relating to poverty. The focus of our analysis is on the remaining four deprivation dimensions, which are as follows:
Basic deprivation: comprising items relating to enforced absence of a meal, clothes, a leisure activity, a holiday, a meal with meat or a vegetarian alternative, adequate home heating, shoes.
Consumption deprivation: comprising three items relating to a PC, an internet connection and a car. It is obviously a rather limited measure and it would be preferable to have a number of additional items.
Health: captured by three items relating to the health of the HRP, namely current reported self-assessed health status, restrictions on current activity and the presence of a chronic illness.
Neighbourhood environment: the quality of the neighbourhood/area environment as reflected in a set of five items comprising reported level of litter, damaged public amenities, pollution, crime/violence/vandalism and noise in the neighbourhood.
The reliability for these dimensions, as indexed by Cronbach’s alpha, ranges from 0.85 for basic deprivation to 0.64 for neighbourhood environment (Whelan and Maître, 2012). Variation in levels of reliability across countries is extremely modest. For each dimension we have used prevalence weighting across the range of countries included in the analysis. This involves giving greater weight to items for which deprivation is less widely experienced by weighting each item by the proportion of households in the overall European sample experiencing the deprivation. In a final step we normalize scores on each of these dimensions so that they have a potential range running from 0 to 1. The former indicates that the household is deprived in relation to none of the items included in the index while the later indicates that they experience deprivation in relation to all of the items.
Our multidimensional analysis of poverty focuses on the four deprivation dimensions described above, together with the ‘at risk of poverty’ indicator identifying those below an income threshold set at 60 percent of median equivalized disposable income in the country in question. For the four deprivation dimensions, there is no natural or general threshold. In conducting multidimensional analysis involving the adjusted head count ratio, interpretation of results is greatly facilitated by choosing comparable thresholds in terms of the number above the cut off. This ensures that the degree of overlap between those deprived on dimensions can range from zero to one. In addition, decomposition of dimensions is an important feature of the approach and interpretation of such analysis is greatly facilitated by this choice. Since the 60 percent of median income at-risk-of-poverty threshold has broad acceptance as an official poverty measure, we have chosen to take it as the benchmark. Weighting for population differences across countries, this income poverty measure identifies 15.7 percent of individuals in the sample as below the income threshold. Therefore we have chosen thresholds for each dimension that come as close as possible to identifying 15.7 percent of individuals as ‘deprived’. 6 This approach minimizes the impact of differences in marginal levels of deprivation across dimensions. We have chosen not to weight dimensions. We define as multi-dimensionally poor those individuals who are above the deprivation-specific threshold on at least two dimensions.
The relationships between deprivation dimensions: censored and uncensored approaches
Here we explore the consequences for the relationships between our selected deprivation dimensions of moving to a censored approach. In Table 1 we show the correlations between each of the dimensions. The uncensored outcomes are above the diagonal and the censored ones below. Focusing first on the former we can see that the highest correlation of 0.395 is between basic and consumption deprivation. Of the remaining correlations, only those relating to the correlation between these dimensions and relative income poverty exceed 0.2. The average correlation is 0.144. Focusing on the uncensored correlations will inevitably lead to an extremely modest estimate of multiple deprivation. Our expectation is that dimensions that may be loosely associated when we consider the population as a whole will be much more closely linked when the comparison with deprivation scores above zero relate only to the multidimensionally poor. The findings in the bottom half of Table 1 confirm these expectations. We find a stronger pattern of correlation between dimensions, reflected in an average correlation of 0.332, which is over double that in the uncensored case.
Censored and uncensored correlation matrices for deprivation dimensions: uncensored above the diagonal and censored below.
Multidimensional poverty levels by country
In Table 2 we show the breakdown by country for the relative income poverty measure, M0 the adjusted head count ratio, H the headcount and I the mean intensity. To facilitate interpretation we have ordered countries by their gross disposable income per capita (GNDH). In column (i) we see the familiar pattern in relation to the at-risk-of-poverty measure, with very modest variation across countries. Somewhat higher levels are observed in the countries with the lowest income levels. Conversely, rates in countries such as the Czech Republic, Slovakia and Slovenia are considerably lower than in a range of countries with higher income levels. The H headcount figures in column (ii), indicate the number above the multidimensional poverty threshold as a consequence of being above the cutoff point on at least two dimensions. In contrast to the at-risk-of-poverty measure, we observe very sharp variation across countries, which is broadly in line with average income levels. The headcount figure ranges from a low of 0.067 in Iceland to a high of 0.592 in Romania. There is a clear tendency for the Scandinavian social democratic countries and the Netherlands to report rates that are lower than might have been expected purely on the basis of their average income levels. By contrast, Greece and Hungary, each of which have experienced particularly difficult recent economic circumstances, exhibit rates somewhat higher than one might have expected from their average incomes.
Multidimensional poverty by country EU-SILC 2009.
Column (iii) focuses on I the average intensity level among those who have been identified as multi-dimensionally poor. The intensity levels are rather similar across countries. There is clearly a relationship between national income levels and intensity with seven of the 11 countries with rates above 0.5 being among the eight lowest income countries. However, outside these countries, variation is extremely modest. The headcount and intensity levels are clearly correlated but variation relating to the former is a great deal more pronounced.
In column (iv) we focus on the adjusted head count ratio. Where no one in the population experiences any of the deprivations it will take on a value of 0 and where every individual experiences deprivation on all items the value will be 1. Our observed range of values goes from 0.030 for Iceland to 0.313 in Romania. The intra-class coefficient (ICC) is 0.108, indicating that just over 10 percent of the total variance is accounted for by between country differences. As with the headcount index, values generally increase as country income levels rise. Once again, values for countries in the social democratic welfare regime are distinctively low. They range from 0.030 in Iceland to 0.060 in the Netherlands and Norway. Countries that show slightly higher values than might be expected on the basis of their income levels are Germany, the UK, Greece and most particularly Hungary. For each of the three lowest income countries the adjusted head count ratio exceeds 0.250. In interpreting these results it is important to remember that a score of 1 would indicate the highly implausible outcome that every individual is above the deprivation threshold on all of the dimensions. An outcome which reaches 20 percent of that level for a European country is clearly one at the extreme end of the continuum.
The adjusted head count ratio measure is a great deal more successful in capturing cross-country variation than the at-risk-of-poverty indicator. While the sharpest differential in the latter case is 2.3 in the former it reaches 10.4. The figures for the adjusted head count ratio index can be contrasted with those for the union and intersection counts set out in columns (v) and (vi). For the former, where all individuals experiencing deprivation on any of the dimensions is counted, the levels range from lows of 0.310 in Iceland and 0.381 in Luxembourg to highs of 0.808 and 0.821 in Bulgaria and Romania, respectively. The figures in relation to the intersection of the dimensions, involving deprivation on all five dimensions, provide a sharp contrast. Here the counts range from close to zero in a large number of countries to 0.012 in Bulgaria and 0.016 in Latvia. The fact that the income poverty variable is defined in relative terms contributes to the extreme nature of these results. However, they are generally consistent with earlier research focusing on multiple deprivation in the EU (Tsakloglou and Papadopoulos, 2002; Whelan et al., 2002).
Decomposition of multidimensional poverty by dimension
One of the advantages of the adjusted head count ratio measure is that it is decomposable in terms of dimensions of deprivation. The overall ratio is equal to the average of the adjusted ratios for the individual dimensions. Similarly, the percentage contribution of a given dimension to overall multidimensional poverty is its weighted ratio divided by the overall ratio. In the present case, since we have applied equal weights, this simply involves dividing the dimension specific ratio by the number of dimensions (thus applying a weight of 0.2) before dividing by the overall ratio. A similar procedure applies in relation to decomposition by socio-economic group with the size of the group substituting for the dimension weight.
In Table 3 we show the decomposition of dimensions broken down by country. It is clear that there is substantial variation across countries in the relative importance of dimensions. In the more affluent countries basic and consumption deprivation play less prominent roles. In only four of the 15 most affluent countries does the figure for basic deprivation rise above 0.20 and in only five does it do so for consumption. In no case is this value exceeded for both dimensions. The combined basic and consumption deprivation rates range from 0.264 in the Netherlands to 0.421 in Germany. In only two countries does it exceed 0.40. For neighbourhood environment the observed rate exceeds 0.20 only for the Netherlands the UK and Italy. For these countries the largest contributors to the adjusted head count ratios are the at-risk-of poverty and health dimensions. For the combined at-risk-of-poverty and health dimensions the rate varies from 0.443 in Germany to 0.539 in Norway.
Decomposition of the adjusted head count social exclusion ratio by dimension by country EU-SILC 2009.
The pattern for the six least affluent countries provides a sharp contrast. The lowest value of the basic deprivation rate of 0.242 is observed for Poland and the highest values of 0.329 and 0.347 for Romania and Bulgaria, respectively. For consumption deprivation the rates range from 0.220 in Hungary to 0.309 in Romania. The combined basic and consumption deprivation rate goes from 0.481 in Latvia to 0.638 in Romania. For all of these countries the contribution of neighbourhood environment is particularly modest and for the three least affluent countries the same is true of the at-risk-of-poverty indicator and health deprivation.
For the remaining countries variation across dimensions is somewhat more variable. The at-risk-of-poverty measure makes a modest contribution in Slovakia and the Czech Republic. In addition, Portugal and Estonia exhibit distinctively low rates of neighbourhood deprivation.
Socio-economic variation in risk levels for multidimensional poverty
At this point we explore the extent to which the impact of social class and age group on the risk of exposure to multidimensional poverty varies across countries. In Table 4 we break down the adjusted head count ratio by an aggregated seven-category version of the European Socio-economic Classification (ESeC) schema for the household reference person (HRP) for each of the countries in our analysis. 7 The class category for which the sharpest degree of variation is observed is farmers, where the range runs from 0.020 in Norway to 0.434 in Romania. Values are generally extremely low in the Scandinavian countries. The ratio rises to between 0.050 to 0.135 for the remaining affluent northern European countries, the Czech Republic, Slovakia and Estonia. Values rise to between 0.159 to 0.248 for Cyprus, Greece and Portugal. Finally the remaining eastern European countries display considerable variation. Hungary, Poland and Lithuania have values close to those of the southern European countries. In contrast, for the least prosperous countries the values range from 0.262 in Latvia to 0.434 in Romania.
Adjusted head count ratio by social class and country.
Note: n/a, not available.
For the remaining categories we observe a similar pattern of class differentiation across countries. Generally the lowest values are observed for the higher professional and managerial group. We also observe a consistent increase in rates moving from the more affluent to the less affluent countries. The rate ranges from 0.007 in Luxembourg to 0.135 in Bulgaria. The next lowest level is observed for the lower professional and managerial class where the rates go from 0.011 in Norway to 0.177 in Bulgaria. The corresponding range for the lower white collar group is from 0.016 in Norway to 0.246 in Bulgaria. For the self-employed group the corresponding figures are 0.032 in Norway to 0.328 in Romania. For the higher working-class group, the observed range goes from 0.033 in Iceland to 0.296, 0.313 and 0.319 respectively for Latvia, Bulgaria and Romania. Finally, the highest adjusted poverty ratio is generally associated with the routine working-class group and those classified as having never worked. The range runs from between 0.038 and 0.074 in Iceland and Norway, respectively, to in excess of 0.330 in Hungary, Latvia, Romania and Bulgaria.
The adjusted head count ratio index clearly fulfils the key requirements of a valid poverty measure in that it varies systematically by social class group within countries and by average income levels across countries. The combined effect is reflected in the fact that the full range of variation for the index runs from 0.007 for the higher professional managerial class in Luxembourg to 0.371 for the routine working-class and never-worked group in Bulgaria – a disparity ratio of 53:1. Social class differences are substantial in every country. The cumulative effects of social class and country produces a situation whereby the most favoured social classes in the least affluent countries exhibit higher poverty rates than the least privileged in the more affluent countries. Thus in Norway and Denmark the value of the adjusted head count ratio is 0.074 and 0.086, respectively, for the routine working-class and never-worked group, while in Latvia and Bulgaria the values for the professional and managerial class are 0.123 and 0.135, respectively. 8
In Table 5 we show the breakdown of the adjusted head count ratio by the age group of the household reference person. Variation across the life course is modest among the more affluent countries. However, there is a tendency for the ratio to be highest for those aged <30 years. For eight of the ten countries with the highest average incomes per capita the disparity ratio involving the comparison of the ≥65 years old group to the <30 years old group does not exceed one. On the other hand, for all 13 lowest income countries the highest level of the adjusted head count ratio is observed for the group aged ≥65 years. In the more affluent countries the lesser importance of basic and consumption deprivation seems to mute the impact of age differences. Where health deprivation comes in combination with basic deprivation it produces a clear pattern of age differentiation. Where it is to a significant extent detached from such deprivation then that is not the case. This may be because the impact of socio-economic deprivation on health is more clearly seen in older age groups.
Mean adjusted head count social exclusion ratio by age group by country EU-SILC 2009.
Multilevel analysis of multidimensional poverty
Since all members of a household are assigned identical values on each of the dimensions it is not possible to conduct an analysis of variation in such outcomes within households. Consequently the multi-level analysis that follows distinguishes between countries and households within countries.
In Table 6 we present a set of hierarchical multilevel regressions with the adjusted head count ratio as the dependent variable. Taking into account such clustering of households within countries allows us to avoid falling prey to the ecological fallacy of drawing erroneous inferences about individual relationships on the basic of aggregate outcomes (Hox, 2010). Column (i) of Table 6 shows the results for the empty model with no independent variables. The ICC capturing the between-country variance as a proportion of the total variance has a value of 0.108. In column (ii) we enter a set of variables relating to the household and household reference person characteristics. These comprise HRP social class, education, marital and parental status, age group and housing tenure. The pattern of results is very much as we would have expected with the ratio being higher for the most disadvantaged educational, class and labour force status, marital and parental status and tenure groups. The inclusion of this set of variables reduces the deviance measured as −2 the log likelihood ratio which is distributed as Chi-squared by 22,543 for 19 degrees of freedom. Taking into account compositional differences in relation to such socio-economic attributes reduces the country variance by 1.9 percent, the individual variance by 11.7 percent and the total variance by 10.6 percent.
Multilevel random intercept model for adjusted head count ratio with household reference person (HRP) characteristics and macro predictors.
Notes: *p<0.05 **p<0.01, ***p<0.001. GNDH, gross disposable income per capita; NS, not significant.
In equation (iii) we explore the impact of adding potentially important macroeconomic influences on multidimensional poverty. In particular, we focus on the log of GNDH and the Gini summary measure of income inequality, with both these variables calculated as deviations from the mean to make later interaction analysis easier to interpret. The values of the Gini variable have also been multiplied by 10 to ease interpretation. The addition of these variables produces a modest reduction in the deviance of 20. The Gini variable is not statistically significant but gross income per capita with a coefficient of −0.152 is highly significant. This model reduces the country variance of the null model by 67.9 percent but has no further effect on the household variance.
In equation (iv) we provide a formal test of the manner in which socio-economic factors interact with gross income per capita. The coefficients for the socio-demographic variables involved in the interactions are their values at the mean of the log of gross income per capita. Looked at another way, the coefficient for the log of gross income per capita is the value where the set of socio-demographic variables take on the reference category or benchmark values. The interaction terms show a consistent pattern of negative coefficients whereby socio-economic disadvantage has a more pronounced effect at lower levels of gross income per capita. Similarly, being in an older age group has a sharper effect in less affluent countries. In other words, cross-country differences in the adjusted head count ratio are significantly higher for the less favoured groups. For example, the coefficient of 0.121 for pre-primary education captures the impact at the mean of the log of gross income per capita. The significant negative interaction of −0.040 indicates that the effect of such education relative to third level education declines as the mean level of gross national income per capita increases and is accentuated at lower levels of affluences. Similarly the significant coefficient of −0.025 for the <30 years old group shows that, at the mean level of log gross income per capita, this group is significantly less likely to be multi-dimensionally poor than the ≥65 years old group. However, the positive interaction coefficient of 0.096 indicates that this negative effect declines as the national income level increases and is correspondingly magnified as it decreases.
Taking into account the manner in which household and HRP socio-economic characteristics interact with national income reduces the deviance figure for equation (iii) by 2,604 for 10 degrees of freedom. Overall the model reduces the between country variance of the empty model by 71 percent, the household variance by 11.7 percent and the total variance by 18.2 percent. The multi-level model analysis confirms that the adjusted head count ratio varies systematically across socio-economic groups and countries but that a fuller understanding of these effects requires that we take into account the manner in which micro and macro factors interact.
Conclusions
This paper has applied to European countries the multidimensional poverty measurement approach recently developed by Alkire and Foster (2011a,b), using data from the EU-SILC 2009 special module on deprivation. Our analysis reveals that the union and intersection approaches to such an exercise produce sharply contrasting results. The former identifies very small numbers as experiencing multidimensional poverty in the better-off countries, while the latter shows a very substantial proportion of the population doing so in every country. Application of the adjusted head count ratio approach developed by Alkire and Foster represents a middle ground between these two extremes. Central to this approach is a censoring of data that counts deprivations only for those fulfilling the dual conditions of being above the deprivation dimension specific threshold and also above the cut-off point for the number of dimensions on which the individual is deprived. Having partitioned individuals in that fashion, the strength of the correlations between the deprivation dimensions is then substantially increased, as is the socio-economic patterning of multidimensional poverty.
We found that this adjusted head count ratio approach identified a non-trivial group as experiencing multidimensional poverty in each of the 28 countries covered, with the size of this group varying with the country’s level of average income per capita. The main source of such variation derives from corresponding variation in the multidimensional head count: while the intensity level is also related to national income, that variation is relatively modest. A decomposition of multidimensional poverty by dimension also revealed that in the less affluent countries basic and consumption deprivation play a more prominent role, while in the richer countries relative income poverty and health were the key factors. The impact of variables such as social class, education and age on multidimensional poverty measured in this way were significantly stronger in low income countries. Thus both the nature of multidimensional poverty and the extent to which it is socially stratified varies by national level of income.
In contrast, the EU decision to frame its 2020 poverty reduction target in terms of the union of at-risk-of-poverty, a (rather extreme) measure of material deprivation, and an indicator of work intensity derives from a series of uneasy compromises with unfortunate consequences. The three dimensions are of varying relevance across the countries of the EU, and the profiles of those it captures as ‘at risk of poverty or social exclusion’ also vary significantly across dimensions (Copeland and Daly, 2012; Nolan and Whelan, 2011; Whelan et al., 2013). The ad hoc manner in which the EU poverty target has been framed serves to highlight the advantages of more structured approaches such as the one employed here. There has been robust debate about the relative merits of an aggregate indicator such as the composite United Nations Development Programme (UNDP) Human Development Index versus the set of Millennium Development Goals, which avoid such aggregation across dimensions, and a similar contrast can be drawn between the composite EU poverty reduction target and the EU’s full suite of social inclusion indicators. Without arbitrating on the relative value of these alternatives, here we have emphasized that where a multidimensional poverty index is constructed, there is much to be gained from adopting an approach with clearly understood axiomatic properties. Doing so allows one to evaluate the consequences of the measurement strategy employed for the levels of multidimensional poverty found, the patterning of such poverty and the associated socio-economic composition and risk profiles, essential in making an informed assessment of the strengths and weaknesses of the particular choices made.
Footnotes
Funding
This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.