Policy Research Working Paper 9120 Assessing Deprivation with an Ordinal Variable Theory and Application to Sanitation Deprivation in Bangladesh Suman Seth Gaston Yalonetzky Development Economics Knowledge and Strategy Team January 2020 Policy Research Working Paper 9120 Abstract The challenges associated with poverty measurement with poverty alleviation programs. To assess the robustness of a cardinal variable have received due attention during the poverty comparisons to alternative choices of parameters, last four decades, but there is a dearth of literature studying various stochastic dominance tests are developed (some of how to meaningfully assess poverty with an ordinal variable. which are themselves novel contributions to the stochastic This paper first proposes a class of simple, intuitive and pol- dominance literature). The empirical illustration docu- icy-relevant poverty measures for ordinal variables. These menting changes in sanitation deprivation in Bangladesh measures are sensitive to depth of deprivations, unlike the showcases the measures’ ability to identify instances in headcount ratio. Moreover, under appropriate restrictions, which overall sanitation deprivation improved while leav- the measures ensure that priority is given to the poorest ing the poorest behind. among the poor while targeting, monitoring and evaluating This paper is a product of the Knowledge and Strategy Team, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at S.Seth@leeds.ac.uk and G.Yalonetzky@leeds.ac.uk. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Assessing deprivation with an ordinal variable: theory and application to sanitation deprivation in Bangladesh Suman Seth∗ Gaston Yalonetzky† 1. Introduction Around four decades ago, in an influential article titled ‘Poverty: An Ordinal Approach to Measurement’, Nobel laureate Amartya Sen proposed an axiomatically derived poverty measure to avoid some shortcomings of the traditionally used headcount ratio (Sen, 1976). Sen’s approach was ordinal in the sense that his poverty measures assigned an ordinal-rank weight to each poor person’s income, an otherwise cardinal variable. Since then, this seminal article has influenced a well-developed literature on poverty measurement involving cardinal variables within an axiomatic framework (Thon, 1979; Clark et al., 1981; Chakravarty, 1983; Foster et al., 1984; Foster and Shorrocks, 1988a,b; Ravallion, 1994; Shorrocks, 1995). ∗ Leeds University Business School, University of Leeds, UK and Oxford Poverty & Human Development Initiative, University of Oxford, UK, S.Seth@leeds.ac.uk. † Leeds University Business School, University of Leeds, UK, G.Yalonetzky@leeds.ac.uk. Distances between the values of cardinally measurable variables are meaningful; whereas 1 ordinal variables merely consist of ordered categories. When numeral scales are as- signed to these categories according to their order, the implied cardinal distances are hard to interpret. Moreover, poverty orderings are inconveniently sensitive to changes in the choice of these scales. Let us consider a simple example with a five-category ordinal vari- able, where the four worst categories reflect deprivations. Suppose the population distribu- tions across the five ordered categories in two countries are A = (0.1, 0.3, 0.2, 0.2, 0.2) and B = (0.2, 0.1, 0.1, 0.2, 0.4), respectively. If we use the numeral scales of (1, 2, 3, 4, 5) for the five categories and use the well-known poverty gap measure with the numeral scale of the fourth category (i.e., the second best category) as the poverty line, then the former society (A) reflects higher poverty than the latter (B).2 However, if we simply change the numeral scales to (1, 5, 6, 7, 8), then the poverty ordering reverses. Yet, the practice of using ordinal variables has been on the rise, in both developed and developing countries alike, due to the recent surge of interest in studying deprivation in non-monetary indicators, which are often ordinal in nature (e.g. access to basic facilities 1 Based on the classification of measurement scales by Stevens (1946), whenever numeral scales are assigned to different ordered categories of an ordinal variable according to the orders or ranks of these categories, any ‘order-preserving’ or monotonic transformation should leave the scale form invariant. See Roberts (1979) for further in-depth discussions. In this paper, by ordinal variables we simply refer to variables with ordered categories, where numeral scales may not have necessarily been assigned to the categories. 2 Let xi be the attainment of individual i and z be the poverty line. Then the poverty gap measure is the average attainment shortfall from the poverty line as a proportion of the latter, attributing a zero shortfall to non-poor people. That is, if N stands for the population size and q stands for the number of poor people, 1 q z −x i then the poverty gap measure (P G) is: P G = N i=1 z (Foster et al., 2013). 2 of different quality).3 Moreover, there may be instances where ordinal categories of an otherwise cardinally measurable variable could have more policy relevance. For example, in some cases we might want to focus on ordered categories of income, nutritional status, or years of education completed, rather than these indicators’ cardinal values. How should poverty be meaningfully assessed with an ordinal variable? One straightforward way may be to dichotomise the population into a group of deprived and a group of non- deprived people, and then use the headcount ratio. However, this index is widely accused of ignoring the depth of deprivations (Foster and Sen, 1997). In our illustration in Section 5, for instance, in Sylhet province of Bangladesh, between 2007 and 2011, the proportion of population with inadequate sanitation facilities went down from around 70% to nearly 63%; whereas, during the same period, the proportion of people with the worst form of sanitation deprivation (‘open defecation’) increased significantly, from around 2% to more than 12% (see Table 3). How can one reasonably capture the depth of deprivations for an ordinal variable? One approach may be to use an aggregate poverty measure that is sensitive to depth of depriva- tions. Alternatively one could simply either consider separately the relative frequencies of the population experiencing deprivation within each ordered category, or equivalently compute and compare headcount ratios (i.e. cumulative frequencies) for each possible poverty line. The latter two approaches seem reasonable especially when the variable under consideration 3 For example, as part of the first Sustainable Development Goals, the United Nations has set the target to not only eradicate extreme monetary poverty, but also to reduce poverty in all its dimensions by 2030. See http://www.un.org/sustainabledevelopment/poverty/ (accessed in April 2017). 3 has only a few deprivation categories and also when the required number of population- comparisons is relatively small. Otherwise they may become cumbersome: even with only four deprivation categories used in our illustration, an analysis of deprivation dynamics across six provinces of Bangladesh over three years would involve comparing a staggering number of 72 data points. Nonetheless, and in order to avoid a multitude of comparisons, one may choose to prioritise and focus on those experiencing the most severe category of deprivation; but this route comes at the expense of ignoring deprivations in other categories. Thus, in this paper, we pursue the former approach of using an aggregate poverty measure that is sensitive to all depths of deprivations within a variable. The challenges associated with measuring well-being and inequality using an ordinal vari- able in an axiomatic framework have received due attention during the last few decades (e.g., Mendelson, 1987; Allison and Foster, 2004; Apouey, 2007; Abul Naga and Yalcin, 2008; Zheng, 2011; Kobus and Milos, 2012; Permanyer and D’Ambrosio, 2015; Kobus, 2015; Lazar and Silber, 2013; Yalonetzky, 2013; Gravel et al., 2015). Yet, when assessing poverty, such efforts have not been sufficiently thorough. Bennett and Hatzimasoura (2011), in a rare attempt, showed that indeed we can measure poverty with ordinal variables sensibly, but implicitly ruled out entire classes of well-suited measures (as shown by Yalonetzky, 2012). Moreover, their assessment of depth-sensitivity was restricted to the ordinal version of Pigou-Dalton transfers, thereby missing many other options including the burgeoning use of Hammond transfers (e.g. see Ebert, 2007; Gravel et al., 2015; Cowell et al., 2017; Gravel et al., 2018; Oui-Yang, 2018).4 4 We refer to the unidimensional context here. The issue of ordinality has certainly been examined thoroughly 4 Our paper contributes theoretically to the poverty measurement literature in three ways. First, we axiomatically characterise a class of ordinal poverty measures under a minimal set of desirable properties. Our class consists of measures that are weighted sums of population proportions in deprivation categories, where these weights are referred to as ordering weights because their values depend on the order of the categories. Our proposed measures are sensitive to the depth of deprivations (unlike the headcount ratio), additively decomposable, and bounded between zero and one. Second, an adequately designed poverty measure should also ensure that policy makers have proper incentive to prioritise those poorer among the poor in the design of poverty alleviation policies so that the poorest are not left behind.5 In a novel attempt, we operationalise the concept of precedence to poorer people by incorporating a new form of degree of poverty aver- sion in the ordinal context, reflecting a prioritarian point of view rather than an egalitarian one. Although grounded on prioritarianism, our new form of poverty aversion encompasses, as limiting cases, both previous attempts at sensitising ordinal poverty indices to the depth of deprivations (e.g., Bennett and Hatzimasoura, 2011; Yalonetzky, 2012) as well as current burgeoning approaches to distributional sensitivity in ordinal frameworks based on Ham- mond transfers (Hammond, 1976; Gravel et al., 2015). We define a range of properties based on this new form of degree of poverty aversion and characterise the corresponding subclasses in the context of multidimensional poverty measurement (Alkire and Foster, 2011; Bossert et al., 2013; Dhongde et al., 2016; Bosmans et al., 2017). However, even in the multidimensional context, ordinal variables are often dichotomised in empirical applications (see, Alkire and Foster, 2011; Bossert et al., 2013; Dhongde et al., 2016), thereby ignoring the depth of deprivations within indicators. 5 Poverty measures may affect the incentives of policy makers during poverty alleviation (Zheng, 1997). 5 of ordinal poverty measures. Within our framework, different degrees of poverty aversion merely require setting different restrictions on the ordering weights, preserving the measures’ additive decomposability property. Third, since each of our classes and subclasses admits a large number of poverty measures, we develop related stochastic dominance conditions whose fulfilment guarantees the robust- ness of poverty comparisons to alternative functional forms and measurement parameters. Although some of these conditions turn out to be the ordinal-variable analogue of existing dominance conditions for cardinal variables (Foster and Shorrocks, 1988b), several others are novel methodological contributions to the literature on stochastic dominance with ordinal variables (to the best of our knowledge). To demonstrate the efficacy of our approach, we present an empirical illustration studying the evolution of sanitation deprivation in Bangladesh using Demographic Health Survey datasets. Interestingly, our measures are able to discern the instances where the improvements in overall sanitation deprivation did not necessarily include the poorest. Furthermore, we apply the stochastic dominance conditions to test the robustness of poverty comparisons over time. The rest of the paper proceeds as follows. After providing the notation, we present and then axiomatically characterise the class of depth sensitive ordinal poverty measures in Section 2. Section 3 introduces the concept of precedence to poorer people, states the properties related to the degrees of poverty aversion, and characterises the subclass of relevant poverty indices. Then section 4 develops stochastic dominance conditions for the characterised class and subclasses of poverty measures. Section 5 provides an empirical illustration analysing sanitation deprivation in Bangladesh. Section 6 concludes. 6 2. A class of depth-sensitive poverty measures for an ordinal variable Suppose, there is a social planner whose objective is to assess a hypothetical society’s poverty in some well-being dimension, which is measured with a set of ordered categories. For instance, self-reported health status may only include response categories, such as ‘good health’, ‘fair health’, ‘poor health’, and ‘very poor health’. Similarly, there are also instances where the ordinal categories of an otherwise cardinal variable, such as the years of schooling completed, have more policy relevance. Formally, suppose, there is a fixed set of S ∈ N \ {1} ordered categories c1 , . . . , cS , where N is the set of positive integers. The ordered categories are such that cs−1 D cs for all s = 2, . . . , S , where D, which reads as “is more deprived than”, is a binary and transitive relation whereby category cs−1 represents a worse-off situation than category cs . Thus, cS is the category reflecting least deprivation and c1 is the state reflecting highest deprivation. Suppose, for example, a society’s well-being is assessed by the education dimension and the observed ordered categories are ‘no education’, ‘primary education’, ‘secondary education’, and ‘higher education’, such that ‘no education’ D ‘primary education’ D ‘secondary education’ D ‘higher education’. Then, c4 = ‘higher education’ and c1 = ‘no education’. We denote the set of all S categories by C = {c1 , c2 , . . . , cS } and the set of all categories excluding the category of least deprivation cS by C−S = C \ {cS }. Each individual in society must experience only one of the S categories. We denote the proportion of population experiencing category cs by ps for all s = 1, . . . , S . Clearly, ps ≥ 0 S for all s and s=1 ps = 1. The proportions of population in the society is summarised by 7 the vector: p = (p1 , . . . , pS ). Note that p is nothing but the discrete probability (or relative frequency) distribution of the society’s population across the S categories. We denote the set of all possible discrete probability distributions over S categories by P. It is customary in poverty measurement to define a poverty threshold for identifying the poor and the non-poor populations (Sen, 1976). Suppose, the social planner decides that category ck for any 1 ≤ k < S and k ∈ N be the poverty threshold, so that people experiencing cat- egories c1 , . . . , ck are identified as poor ; whereas, people experiencing categories ck+1 , . . . , cS are identified as non-poor. We assume that at least one category reflects the absence of poverty, as this restriction is both intrinsically reasonable and is required for stating certain properties in Section 2.1. When k = 1, only category c1 reflects poverty and, in this case, p1 is the proportion of the population identified as poor. For any ck ∈ C−S , we denote the k proportion of poor population, also known as the headcount ratio, by H (p, ck ) = s=1 ps . We define a poverty measure P (p, ck ) as P : P × C−S → R+ . In words, a poverty measure is a mapping from the set of probability distributions and the set of poverty thresholds to the real line (note that the set of categories remains fixed). 2.1. Desirable properties Now, as in many other areas of well-being measurement, the social planner needs to make certain assumptions to perform interpersonal comparisons and meaningfully aggregate the information available on the ordered deprivation categories in order to obtain a numerical poverty measure. We present these assumptions in terms of properties or axioms. By relying on discrete probability distributions, we already implicitly assume that poverty measures 8 satisfy anonymity and population principle. Anonymity requires that merely shuffling the individual deprivation levels within a society should not alter the society’s poverty level; whereas, the population principle requires that a mere duplication of the number of indi- viduals experiencing each deprivation level within a society should not affect the society’s poverty level. The population principle allows comparing poverty levels of societies with dif- ferent population sizes. The social planner makes the following four additional assumptions. The first property is ordinal monotonicity, which requires that if the living standard of a poor person improves so that the person moves to a category of less deprivation, then societal poverty should be lower. Formally, the property requires that if a poor person or a group of poor people moves from a category ct reflecting poverty (i.e. t ≤ k ) to a less deprived category cu (i.e. ct D cu ), while the deprivation levels of everybody else in the society remain unchanged, then poverty should fall. In terms of probability distributions, we set the requirement that if a fraction of the poor population is moved from category ct to category cu , while the proportion of population in other categories remain unchanged, then poverty should fall: Ordinal Monotonicity (OMN) For any p, q ∈ P and for any ck ∈ C−S , if qt < pt for some t ≤ k and t < u but ps = qs ∀s = {t, u}, then P (q, ck ) < P (p, ck ). The second property is single-category deprivation. The property requires that whenever there is only one category reflecting poverty (i.e. c1 ), then the poverty measure should be equal to the headcount ratio H (p, c1 ) = p1 . In other words, we assume that whenever there is only one category reflecting poverty and the others reflect an absence of poverty, then 9 the headcount ratio becomes a sufficient statistic for the assessment of poverty. In fact, in this situation, any functional transformation of the headcount ratio would not add any meaningful information to the poverty assessment while being inferior in terms of intuitive interpretation. Single-Category Deprivation (SCD) For any p ∈ P and c1 ∈ C−S , P (p, c1 ) = p1 . The third property, focus, is essential for a poverty measure. It requires that, ceteris paribus, a change in a non-poor person’s situation should not alter societal poverty evaluation as long as the non-poor person remains in that status. In terms of probability distributions, we set the requirement that as long as the proportion of poor population within each of the k categories reflecting poverty remains unchanged, the level of poverty should be the same. Note that the proportions of non-poor people may remain unchanged or may be different across the S − k categories not reflecting poverty, but this should not matter for poverty evaluation. Focus (FOC) For any p, q ∈ P and for any ck ∈ C−S , if qs = ps ∀s ≤ k , then P (q, ck ) = P (p, ck ). Finally, the social planner may be interested in exploring the relationship between the overall poverty evaluation and the subgroup poverty evaluation, where population subgroups can be geographical regions, ethnic groups, etc. Suppose society is partitioned into M ∈ N\{1} mutually exclusive and collectively exhaustive population subgroups. We denote the pop- M ulation share in subgroup m by πm , such that πm ≥ 0 ∀m = 1, . . . , M and m=1 πm = 1. 10 We further denote the probability distribution across S categories within subgroup m by M pm = (pm m 1 , . . . , pS ) ∈ P for every m = 1, . . . , M , such that p = m=1 πm pm . The final property, subgroup decomposability, requires overall societal poverty to be expressible as a population-weighted average of subgroup poverty levels: Subgroup Decomposability (SUD) For any M ∈ N/{1}, for any p ∈ P such that p = M M m=1 πm pm where (i) pm ∈ P ∀m = 1, . . . , M , (ii) πm ≥ 0, and (iii) m=1 πm = 1, and for any ck ∈ C−S , M P (p, ck ) = πm P (pm , ck ). m=1 2.2. The class of poverty measures The four properties outlined in Section 2.1 lead to the class of poverty measures that we present in Theorem 2.1: Theorem 2.1 A poverty measure P satisfies properties OMN, SCD, FOC and SUD if and only if S P (p, ck ) = ps ω s (2.1) s=1 where ω1 = 1, ωs−1 > ωs > 0 for all s = 2, ..., k whenever k ≥ 2, and ωs = 0 for all s > k . Proof. See Appendix A1. We denote our class of poverty measures in Theorem 2.1 by P . A poverty measure in our proposed class is a weighted sum of the population proportions in p, where the weights (i.e. ωs ’s) are non-negative for all categories, strictly positive for the deprived categories, 11 and unity for the most deprived category. We refer to weights ωs ’s as ordering weights and to ω = (ω1 , . . . , ωS ) as the ordering weighting vector.6 The ordering weights increase with deprived categories representing higher levels of deprivation. In practice, the ordering weights may take various forms. For example, Bennett and Hatzimasoura (2011) make the ordering weight for each deprivation category depend on the latter’s relative deprivation rank. Category s is assigned an ordering weight equal to ωs = [(k − s + 1)/k ]θ for all s = 1, . . . , k and for some θ > 0. Thus, the least deprived category ck receives an ordering weight of ωk = 1/k θ ; whereas, the most deprived category c1 receives an ordering weight of ω1 = 1. The class of poverty measures in Equation 2.1 bears several policy-relevant features. First, an appealing feature of the measures in our class is their simplicity. Second, unlike the headcount ratio, the poverty measures in our class are sensitive to the depth of deprivations as they assign larger weights to the more deprived categories. Thus, unlike the headcount ratio, the proposed measures are sensitive to changes in deprivation status among the poor even when they do not become non-poor owing to those changes. Third, the proposed poverty measures are additively decomposable, which has two crucial policy implications. One is that the society’s overall poverty measure may be expressed as a population-weighted average of the population subgroups’ poverty measures, whenever the entire population is di- 6 Kobus and Milos (2012, Theorem 3) also showed that a subgroup decomposable inequality measure for ordinal variables that is sensitive to spreads away from the median is a monotonic transformation of the weighted sum of population proportions. However, our ordering weights are significantly different both in terms of their restrictions and interpretation vis-a-vis those involved in ordinal inequality measurement. Moreover, our SCD property (see below) imposes a more stringent restriction on the permissible functional transformations. 12 vided into mutually exclusive and collectively exhaustive population subgroups. The other is that additively decomposable measures are convenient for cross-sectional and inter-temporal econometric analysis as well as impact evaluation exercises. Fourth, the poverty measures are conveniently normalised between zero and one. They are equal to zero only in a soci- ety where nobody is poor; whereas, they are equal to one only whenever everybody in the society experiences the worst possible deprivation category c1 . Note that the normalisation behaviour of our measures between zero and one, especially owing to the restrictions on ω ’s, is not an axiomatic assumption, but a logical conclusion from the foundational properties. Fifth, the poverty measure boils down to the headcount ratio either when the poverty thresh- old is represented by the most deprived category or whenever the underlying ordinal variable has merely two categories. Each poverty measure in class P may also have the following alternative interpretation: when a policy maker only observes the population’s relative frequencies across ordered deprivation categories, then, based on the four assumptions, the policy maker assigns particular depri- vation values in the form of ωs ’s to these individuals. Each poverty measure in our proposed class is an average of these assigned deprivation values.7 3. Precedence to the poorer among the poor Poverty alleviation is a gradual process, where it is imperative to ensure that the poorest of the poor are not left behind. Although all poverty measures in Equation (2.1) are sensitive to 7 For an alternative version of the axiomatic characterisation of the measures in class P using anonymity and population principle alongside the aforementioned properties, see Seth and Yalonetzky (2018). 13 the depth of deprivations, not all of them ensure that the poorest among the poor population receive precedence over the less poor population during a poverty alleviation process. For that purpose, we introduce an intuitive concept of giving precedence to poorer people in the ordinal framework, in tune with the prioritarian view which holds that ‘benefiting people matters more the worse off these people are’ (Parfit, 1997, p. 213).8 Furthermore, our notion of precedence is presented in a general framework where the degree of precedence to poorer people can vary between a minimum and a maximum.9 Let us introduce the concept using the example in Figure 1. Suppose, there are ten ordered deprivation categories, where c1 is the most deprived category and c10 is the least deprived category. Denoted by black circles, categories c1 , . . . , c7 reflect poverty; whereas the other three categories, denoted by gray circles, do not. Thus, c7 is the poverty threshold category, which is highlighted by a large black circle in each distribution. The original distribution is p, where each individual experiences one of the ten categories. Suppose the policy maker has the following two competing options: either (a) obtain distri- bution p from p by assisting a fraction ∈ (0, 1) of poor people to move from category c1 to category c2 ; or (b) obtain distribution q from p by assisting a fraction of poor people to move from category c2 to category c3 . Which option should lead to a larger reduction in poverty? One way of giving precedence to poorer people is to require that the move from 8 For an application of the prioritarian concept to the multidimensional context, see Bosmans et al. (2017). For further discussion, comparing the prioritarian view with the egalitarian view see Fleurbaey (2015, p. 208). 9 The concept is analogous to the degree of poverty or inequality aversion in the monetary poverty measure- ment literature (Clark et al., 1981; Chakravarty, 1983; Foster et al., 1984), but not technically identical. 14 Figure 1: Precedence to poorer and the degree of precedence p c10 c9 c8 ck=7 c6 c5 c4 c3 c2 c1 p c10 c9 c8 ck=7 c6 c5 c4 c3 c2 c1 q c10 c9 c8 ck=7 c6 c5 c4 c3 c2 c1 q c10 c9 c8 ck=7 c6 c5 c4 c3 c2 c1 q c10 c9 c8 ck=7 c6 c5 c4 c3 c2 c1 p to p should lead to a larger reduction in poverty than the move from p to q. It is a minimal criterion for giving precedence to poorer people which we call minimal precedence to poorer people (PRE-M). This property requires that, ceteris paribus, moving a fraction of poorer people to an adjacent less deprived category leads to a larger reduction in poverty than moving a fraction of less poor people to a respectively adjacent less deprived category.10 The PRE-M property presents a minimal criterion for giving precedence to poorer people. Yet what happens when the policy maker faces the alternatives of improving the situation of a fraction of poorer people by one category and improving the situation of a similar fraction of less poor people by several categories? To ensure that the policy maker still chooses to improve the situation of the poorer people in these cases, we introduce the property of greatest precedence to poorer people (PRE-G). This property requires that, ceteris paribus, moving a 10 We have defined only the strict versions of these properties, requiring poverty to be strictly lower in the aftermath of specific pro-poorest distributional change. Consequently the ensuing results impose strict inequality restrictions on weights. However, these strict restrictions may be relaxed with alternative versions if the latter only require that poverty does not rise due to the same pro-poorest distributional change. 15 fraction of poorer people to a less deprived category leads to a larger reduction in poverty than moving a fraction of less poor people to any number of less deprived categories. Note here that the improvement is not restricted to a particular number of adjacent categories. For example, if PRE-G held, a move from p to p in Figure 1 should lead to a larger reduction in poverty than even a move from p to q . Conceptually, the PRE-G property is analogous to the notion of a Hammond transfer (Ham- mond, 1976; Gravel et al., 2015), which essentially involves, simultaneously, an improvement in a poor person’s situation and a deterioration of a less poor person’s situation, such that their deprivation ranks are not reversed (in the case of poverty measurement). Also, an ordinal poverty measure satisfying property PRE-G also satisfies property PRE-M, but the reverse is not true. A policy maker supporting property PRE-G over property PRE-M should be considered more poverty averse. We could also consider intermediate forms of preference between the minimal (PRE-M) and the greatest (PRE-G) forms of precedence. For example, instead of the greatest forms of precedence, the policymaker may be satisfied with requiring, that, ceteris paribus, moving a fraction of poorer people to an adjacent less deprived category leads to a larger reduction in poverty than moving a fraction of less poor people to, say, two adjacent less deprived categories. In such case, a move from p to p in Figure 1 should lead to a larger reduction in poverty than a move from p to q . We refer to this case as precedence to poorer people of order two (PRE-2). Likewise, we may obtain the policy maker’s preferred degree α of giving precedence to poorer people. Thus, we introduce the general property of precedence to poorer people of order α, 16 which requires that, ceteris paribus, moving a fraction of poorer people to an adjacent less deprived category leads to a larger reduction in poverty than moving a fraction of less poor people up to an α (≥ 1) number of adjacent less deprived categories. A formal general statement of the property, which includes PRE-M and PRE-G as limiting cases, is the following: Precedence to Poorer People of Order α (PRE-α) For any p, p , q ∈ P, for any k ≥ 2, for any ck ∈ C−S , and for some α ∈ N such that 1 ≤ α ≤ k − 1, for some s < t ≤ k < S and for some ∈ (0, 1), if (i) p is obtained from p such that ps = ps − while pu = pu ∀u = {s, s + 1}, and (ii) q is obtained from p such that qt = pt − while qu = pu ∀u = {t, min{t + α, S }}, then P (p , ck ) < P (q , ck ). Note that PRE-1 is essentially the PRE-M property. This is the case where the social planner is least poverty averse. As the value of α increases, the social planner’s poverty aversion rises. In this framework, the social planner’s poverty aversion is highest at α = k − 1. We shall subsequently show that PRE-α for α = k − 1 leads to the same subclass of ordinal poverty measures as the PRE-G property. The PRE-α property imposes further restrictions on the class of measures in Theorem 2.1. In Theorem 3.1, we present the subclass of measures Pα that satisfy the general PRE-α property: Theorem 3.1 For any k ≥ 2 and for some α ∈ N such that 1 ≤ α ≤ k − 1, a poverty measure P ∈ P (Theorem 2.1) satisfies property PRE-α if and only if a. ωs−1 − ωs > ωs − ωs+α ∀s = 2, . . . , k − α and ωs−1 > 2ωs ∀s = k − α + 1, . . . , k whenever α ≤ k − 2. 17 b. ωs−1 > 2ωs ∀s = 2, . . . , k whenever α = k − 1. Proof. See Appendix A2. In a novel effort, Theorem 3.1 presents various subclasses of indices based on the degree of poverty aversion α, which we denote as Pα . In order to give precedence to poorer people, additional restrictions must be imposed on the ordering weights. Corollary 3.1 presents the limiting case of P1 , featuring the least poverty averse social planner: Corollary 3.1 For any k ≥ 2, a poverty measure P ∈ P (Theorem 2.1) satisfies property PRE-M (i.e. PRE-1) if and only if ωs−1 − ωs > ωs − ωs+1 ∀s = 2, ..., k − 1 and ωk−1 > 2ωk . Proof. The result follows directly from Theorem 3.1 by setting α = 1. To give precedence to poorer people in the spirit of property PRE-M, the ordering weights must be such that the difference ωs−1 − ωs is larger than the subsequent difference ωs − ωs+1 , in addition to the restrictions imposed by Theorem 2.1. Suppose, we summarise the ordering weights by: ω = (ω1 , . . . , ωS ). Let us consider an example involving five categories and two ordering weight vectors: ω = (1, 0.8, 0.5, 0, 0) and ω = (1, 0.5, 0.2, 0, 0), where k = 3. The ordering weights in ω fulfill all properties presented in Theorem 2.1, but the largest reduction in poverty is obtained whenever a poor person moves from the least poor category to the adjacent non-poor category. By contrast, ordering weights in ω require that the largest reduction in poverty be obtained whenever a poor person moves from the poorest category to the adjacent second poorest category. Thus, unlike the ordering weights in ω , the ordering weights in ω make sure that poorer people receive precedence. 18 Next we present the subclass of poverty measures satisfying property PRE-G, i.e. PG : Proposition 3.1 For any k ≥ 2, a poverty measure P ∈ P (Theorem 2.1) satisfies property PRE-G if and only if ωs−1 > 2ωs ∀s = 2, . . . , k . Proof. See Appendix A3. The additional restriction on the ordering weights in Proposition 3.1 effectively prioritises the improvement in a poorer person’s situation over improvement of any extent in a less poor person’s situation. Let us consider an example involving five categories and two ordering weight vectors: ω = (1, 0.6, 0.3, 0.1, 0) and ω = (1, 0.48, 0.23, 0.1, 0), where k = 4. Clearly, both sets of weights in ω and ω satisfy the restriction in Corollary 3.1 that ωs−1 − ωs > ωs − ωs+1 for all s = 2, . . . , k . However, the ordering weights in ω do not satisfy the restriction in Proposition 3.1, since ω1 < 2ω2 ; whereas the ordering weights in ω do satisfy the restriction in Proposition 3.1 as ωs−1 > 2ωs for all s = 2, . . . , k . An interesting feature of the set of weights satisfying property PRE-G is that, for k ≥ 3, any deprivation category up to the third least detrimental deprivation category (i.e. k − 2), should receive a weight greater than the sum of weights assigned to all categories reflecting k lesser deprivation, i.e. ωs > =s+1 ω for all s = 1, . . . , k − 2. Finally, it is worth pointing out that, remarkably, the subclasses PG (Proposition 3.1) and Pk−1 (Theorem 3.1 when α = k − 1) are identical; even though the distributional changes involved in the PRE-α property are only specific cases of those involved in PRE-G. Besides being of interest in itself, this perfect overlap between the subclasses of indices will prove useful in Section 4 because by deriving the dominance conditions for the subclasses Pα , we 19 will also obtain the relevant dominance conditions for subclass PG .11 4. Poverty dominance conditions Stochastic dominance conditions come in handy whenever we want to ascertain the robust- ness of a poverty ranking of distributions to alternative reasonable comparison criteria, e.g. selection of poverty lines, choice of different functional forms, etc. (Atkinson, 1987; Foster and Shorrocks, 1988b; Fields, 2001). Moreover, often stochastic dominance conditions reduce an intractable problem of probing the robustness of a comparison across an infinite domain of alternative criteria to a finite set of distributional tests (Levy, 2006). In Sections 2 and 3, we introduced the class of poverty measures P and its subclasses Pα . The main parameters for these measures are the set of ordering weights {ω1 , . . . , ωk } and the poverty threshold category ck . It is thus natural to inquiry into the circumstances under which poverty com- parisons based on ordinal variables are robust to alternative ordering weights as well as alternative poverty threshold categories. In this section, first we introduce the first-order dominance conditions relevant to P , followed by the second-order dominance conditions for Pα for all α. 11 Noteworthy is the expected resemblance between the weighting restriction identified in Proposition 3.1 and that in the class of welfare functions for ordinal variables derived in Gravel et al. (2015, Lemma 2). The latter characterises welfare functions that increase both when someone moves to a better category (so-called increments) and in the aftermath of Hammond transfers. Setting αk = 0 and changing the inequality sign in Gravel et al. (2015, Lemma 2) to interpret their α functions (not to be confused with our α parameter) as ordering weights for poverty measurement yields the weight restriction in Proposition 3.1. 20 4.1. Poverty dominance conditions for all measures in P Theorem 4.1 provides poverty dominance conditions that are relevant to all measures in class P , but for a given poverty threshold category ck ∈ C−S : Theorem 4.1 For any p, q ∈ P and a given ck ∈ C−S , P (p, ck ) < P (q, ck ) ∀P ∈ P if and s only if =1 (p − q ) ≤ 0 ∀s ≤ k with at least one strict inequality. Proof. See Appendix A4. Once a particular poverty threshold category ck ∈ C−S is chosen, Theorem 4.1 states that poverty in distribution p ∈ P is strictly lower than that in distribution q ∈ P for all measures in P if and only if all partial sums of the probabilities up to category ck in p are nowhere higher and at least once strictly lower than the respective partial sums in q. The result in Theorem 4.1 may also be presented in terms of the headcount ratio H (p, ck ) as follows: the poverty comparison for a particular poverty threshold category ck is robust to all poverty measures in P if and only if H (p, cs ) ≤ H (q, cs ) for all s ≤ k and H (p, cs ) < H (q, cs ) for at least one s ≤ k . Corollary 4.1 extends the first-order dominance condition obtained in Theorem 4.1 to any measure P ∈ P and all poverty threshold categories in C−S : Corollary 4.1 For any p, q ∈ P and all ck ∈ C−S , P (p, ck ) < P (q, ck ) for any P ∈ P and s for all ck ∈ C−S if and only if =1 (p − q ) ≤ 0 ∀s = 2, . . . , S − 1 and p1 < q1 . Proof. The sufficiency part is straightforward and follows from Equation A1. We prove the necessary condition as follows. First, consider k = 1. Then, P (p, c1 ) < P (q, c1 ) only if 21 s s p1 < q1 . Subsequently, the requirement that =1 p ≤ =1 q ∀s = 2, . . . , S − 1 follows from Theorem 4.1. Interestingly, in terms of headcount ratios, poverty in distribution p is lower than poverty in distribution q for any P ∈ P and for all possible poverty threshold categories if and only if H (p, cs ) ≤ H (q, cs ) for all s ≤ k and H (p, c1 ) < H (q, c1 ). Thus, the results in Theorem 4.1 and Corollary 4.1 are the ordinal versions of the headcount-ratio orderings for continuous variables derived by Foster and Shorrocks (1988b). 4.2. Poverty dominance conditions for all measures in Pα We now present poverty dominance conditions that are relevant to all measures in subclass Pα for some α ∈ [1, k − 1]. We refer to the poverty dominance for a particular value of α as PRE-α dominance. For the presentation of our results in this section, we define the following additional notation. First, by b for any b ∈ R++ , we denote the largest possible non-negative integer that is not greater than b (for example, if b = 5.2 then b = 5). r Secondly, for some k ≥ 2, we denote the rth ordinal weighting vector by ω r = (ω1 r , . . . , ωS ) for all r = 1, . . . , k . In Theorem 4.2, we show that one needs to evaluate poverty orderings across two distributions at k distinct ordinal weighting vectors to test PRE-α dominance r and we explicitly derive the weights ωs for all s = 1, . . . , S and for all r = 1, . . . , k . Theorem 4.2 presents the PRE-α dominance conditions for a given poverty threshold category ck ∈ C−S and k ≥ 2: Theorem 4.2 For any p, q ∈ P, for some k ≥ 2, and for some α ∈ [1, k − 1] ⊆ N, P (p, ck ) < 22 P (q, ck ) ∀P ∈ Pα for a given ck ∈ C−S if and only if, with at least one strict inequality, S r s=1ωs (ps − qs ) ≤ 0 ∀r = 1, . . . , k such that:       0 for s > r and r = 1, . . . , k         1 for s = 1 and r = 1, . . . , k    21−s      for s = 2, . . . , r and r = 2, . . . , α + 1  2r−s  r ωs = for s = r − α, . . . , r and r = α + 2, . . . , k ,  r (−1)j r−αj −1 r−(α+1)j −1 2   j      j =0   s (−1)j r−sj−αj 2r−s−(α+1)j      j =0 for s = 2, . . . , r − α − 1 and r = α + 3, . . . , k     r (−1)j r−αj −1 r−(α+1)j −1     j 2 j =0 (4.1) r−1 r−s where r = and s = . α+1 α+1 Proof. See Appendix A5. For a given ck ∈ C−S and for a given α ∈ [1, k − 1], according to Theorem 4.2, poverty in distribution p is lower than that in distribution q for all P ∈ Pα (i.e., p PRE-α dominates S r S r q) if and only if s=1 ωs ps ≤ s=1 ωs qs ∀r = 1, . . . , k with at least one strict inequality, r where the values for all ωs ’s are determined by Equation 4.1. The first condition within r Equation 4.1 requires that ωs = 0 whenever s > r; whereas, the second condition requires, r based on Theorem 2.1, that ω1 = 1 for all r = 1, . . . , k . For example, for k = 3, ω 1 = 1 1 1 (ω1 , ω2 , ω3 ) = (1, 0, 0). The third condition within Equation 4.1 requires that for values of r r ranging between 2 and α + 1 for a given value of α, ωs = 21−s for all s = 1, . . . , r, due to the restriction ωs−1 > 2ωs ∀s = 2, . . . , r whenever α = r − 1 or r = α + 1 as in part b. of Theorem 3.1. The fourth and fifth conditions within Equation 4.1 follow from the second 23 condition and the first condition, respectively, of part a. of Theorem 3.1. For any chosen value of α, k different conditions must be tested for concluding dominance in a poverty comparison.12 Let us provide an example with S = 6 and k = 5. In this k case, α may take four different values: α = 1, 2, 3, 4. Moreover, since S > k , ω6 = 0 for 6 r 6 r all k . For each value of α, the condition s=1 ωs ps ≤ s=1 ωs qs , with one strict inequality, must be satisfied for the five ordinal weighting vectors ω r corresponding to r = 1, . . . , 5, as presented in Table 1. Each column of the table presents the five ordering weighting vectors for a given value of α; whereas, the k th row of the table reports the k th ordering ¯ are also reported weights vector for different values of α. The corresponding values of r where they are relevant. For instance, if α = 2, then poverty comparisons should be checked at the following five ordering weighting vectors: ω 1 = (1, 0, 0, 0, 0, 0), ω 2 = 1, 1 2 , 0, 0, 0, 0 , ω 3 = 1, 1 4 2 1 , 1 , 0, 0, 0 , ω 4 = 1, 7 2 4 7 4 2 1 , 7 , 7 , 0, 0 , and ω 5 = 1, 12 , 12 , 12 , 12 , 0 . We also provide simplified versions of the dominance conditions presented in Theorem 4.2 for the two extreme cases of minimal precedence (i.e. α = 1) as in Corollary 3.1 and of greatest precedence (i.e. α = k − 1) as in Proposition 3.1. Whenever α = 1, we must test k restrictions on the linear combinations of partial sums of probabilities as in Corollary 4.2: Corollary 4.2 For any p, q ∈ P and for some k ≥ 2, P (p, ck ) < P (q, ck ) ∀P ∈ P1 for s a given ck ∈ C−S if and only if =1 j =1 (pj − qj ) ≤ 0 ∀s ≤ k with at least one strict inequality. 12 This approach to dominance testing is analogous to that used in robustness tests for composite indices with respect to alternative weights (see, Seth and McGillivray, 2018; Foster et al., 2012). 24 Table 1: The ordering weighting vectors for conducting dominance tests for different α’s when S = 6 and k = 5 α=1 α=2 α=3 α=4 ω1 (1, 0, 0, 0, 0, 0) (1, 0, 0, 0, 0, 0) (1, 0, 0, 0, 0, 0) (1, 0, 0, 0, 0, 0) 1 1 1 1 ω2 1, , 0, 0, 0, 0 1, , 0, 0, 0, 0 1, , 0, 0, 0, 0 1, , 0, 0, 0, 0 2 2 2 2 2 1 1, , , 0, 0, 0 ; 1 1 1 1 1 1 ω3 3 3 1, , , 0, 0, 0 1, , , 0, 0, 0 1, , , 0, 0, 0 2 4 2 4 2 4 ¯= 1 r 3 2 1 4 2 1 1, , , , 0, 0 ; 1, , , , 0, 0 ; 1 1 1 1 1 1 ω4 4 4 4 7 7 7 1, , , , 0, 0 1, , , , 0, 0 2 4 8 2 4 8 ¯= 1 r ¯= 1 r 4 3 2 1 7 4 2 1 8 4 2 1 1, , , , , 0 ; 1, , , , , 0 ; 1, , , , ,0 ; 1 1 1 1 ω5 5 5 5 5 12 12 12 12 15 15 15 15 1, , , , , 0 2 4 8 16 ¯= 2 r ¯= 1 r ¯= 1 r Proof. See Appendix A6. The dominance condition presented in Corollary 4.2 is the ordinal version of the “P2 ” poverty ordering as in Foster and Shorrocks (1988b). Corollary 4.3, on the other hand, presents the dominance conditions for greatest precedence (i.e. α = k − 1): Corollary 4.3 For any p, q ∈ P and for some k ≥ 2, P (p, ck ) < P (q, ck ) ∀P ∈ PG for a s given ck ∈ C−S if and only if =1 21− (p − q ) ≤ 0 ∀s ≤ k with at least one strict inequality. Proof. It is straightforward to verify from Theorem 4.2 by setting α = k − 1. So far, we have presented the dominance conditions for a given ck . Finally, Corollary 4.4 provides the second-order dominance conditions relevant to any measure P ∈ Pα for some α ∈ [1, k − 1], but for all ck ∈ C−S such that k ≥ 2: Corollary 4.4 For any p, q ∈ P, for some k ≥ 2, and for some α ∈ [1, k − 1] ⊆ N, P (p, ck ) < P (q, ck ) ∀P ∈ Pα for all ck ∈ C−S if and only if (a) p1 ≤ q1 and 2p1 + p2 ≤ 2q1 + q2 25 S r with at least one strict inequality and (b) s=1 ωs (ps − qs ) ≤ 0 ∀r = 3, . . . , S − 1, where r ω r = (ω1 r , . . . , ωS ) ∀r = 3, . . . , S − 1 are obtained from Equation 4.1 of Theorem 4.2. Proof. First consider the case when k = 2. Then from Equation 4.1 of Theorem 4.2, it follows that p1 ≤ q1 and 2p1 + p2 ≤ 2q1 + q2 , with at least one strict inequality. Therefore, conditional (a) is both necessary and sufficient. Whenever, k ≥ 3, part (b) follows from Theorem 4.2 by setting k = S − 1. In summary, the robustness of poverty comparisons for various classes and subclasses of ordinal measures introduced in Sections 2.1 and 3 can be assessed with a battery of dominance tests based on the theorems and corollaries presented in this section. 5. Empirical illustration: Sanitation deprivation in Bangladesh We now present an empirical illustration in order to showcase the efficacy of our proposed measurement method. In the current global development context, both the United Nations through the Sustainable Development Goals13 and the World Bank through their Report of the Commission on Global Poverty (World Bank, 2017) have acknowledged the need for assessing, monitoring, and alleviating poverty in multiple dimensions besides the monetary dimension. In practice, most non-income dimensions are assessed by ordinal variables. In this section, we show how our measurement tools may be applied to analyse inter-temporal sanitation deprivation in Bangladesh. For our analysis, we use the nationally representative Demographic Health Survey (DHS) 13 Available at https://sustainabledevelopment.un.org/sdgs. 26 datasets of Bangladesh for the years 2007, 2011, and 2014. We compute the point estimates and standard errors incorporating the sampling weights as well as respecting the survey design.14 Excluding the non-usual residents, we were able to consider the information on 50,215 individuals from 10,398 households in the 2007 survey, 79,483 individuals from 17,139 households in the 2011 survey, and 77,680 individuals from 17,299 households in the 2014 survey. Table 2: The five ordered categories of access to sanitation facilities Category Description Open defecation Human faeces disposed of in fields, forests, bushes, open bodies of water, beaches or open spaces or disposed of with solid waste Unimproved Pit latrines without a slab or platform, hanging latrines and bucket latrines Limited Sanitation facilities of an otherwise acceptable type shared between two or more house Basic unsafe A basic sanitation facility which is not shared with other households, but excreta ar disposed safely, such as flushed but not disposed to piped sewer system, septic tank o latrine Improved Sanitation facility which is not shared with other households and where excreta are s disposed in situ or treated off-site and includes flush/pour flush to piped sewer system, s tank or pit latrine, ventilated improved pit latrine, composting toilet or pit latrine w slab One target of the United Nations’ sixth Sustainable Development Goal (whose aim is to ‘ensure availability and sustainable management of water and sanitation for all’) is ‘by 2030, [to] achieve access to adequate and equitable sanitation and hygiene for all and end open defecation.’ In order to hit the target, the Joint Monitoring Programme (JMP) of the World Health Organisation and the UNICEF proposes using ‘a service ladder approach to benchmark and track progress across countries at different stages of development”, building 14 See NIPORT et al. (2009, 2013, 2016) for details about the survey design. 27 on the existing datasets.15 We pursue this service ladder approach and apply our ordinal poverty measures to study the improvement in sanitation deprivation in Bangladesh. We classify households’ access to sanitation in the five ordered categories presented in Table 2. The five categories are ordered as ‘open defecation’ D ‘unimproved’ D ‘limited’ D ‘basic unsafe’ D ‘improved’. We consider all persons living in a household deprived in access to sanitation if the household experiences any category other than ‘improved’. Table 3: Change in population distribution across sanitation categories in Bangladesh Bangladesh Dhaka Rajshahi Sylhet 2007 2011 2014 2007 2011 2014 2007 2011 2014 2007 2011 2014 Open defecation 7.5 4.2 3.3 7.5 4.0 2.2 13.8 3.9 3.2 2.1 12.5 9.4 (0.8) (0.3) (0.5) (1.4) (0.7) (0.8) (2.3) (0.8) (0.9) (0.4) (1.5) (1.3) Unimproved 47.1 38.3 25.7 44.3 35.9 22.6 45.3 36.4 28.3 57.2 34.3 23.0 (1.1) (0.9) (1.3) (1.9) (1.7) (2.8) (2.6) (3.1) (2.7) (3.3) (2.0) (2.6) Limited 13.4 16.7 20.9 14.4 18.0 26.1 14.7 20.7 20.2 10.1 17.7 22.3 (0.5) (0.6) (0.8) (1.1) (1.5) (2.0) (1.2) (1.4) (1.3) (1.9) (0.9) (1.8) Basic 3.5 4.3 2.3 8.6 10.5 5.4 0.2 0.2 0.3 0.6 0.1 0.3 (0.4) (0.5) (0.4) (1.0) (1.6) (1.0) (0.1) (0.1) (0.1) (0.2) (0.1) (0.2) Improved 28.5 36.6 47.8 25.2 31.6 43.7 26.0 38.8 48.0 30.1 35.4 45.0 (1.0) (0.9) (1.1) (1.9) (1.9) (2.5) (1.9) (2.3) (2.3) (2.3) (1.9) (1.7) Sources: Authors’ own computations. Standard errors are reported in parentheses. Table 3 shows how the estimated population shares in different deprivation categories have evolved over time in Bangladesh. Clearly, the estimated percentage in the ‘improved’ cate- gory has gradually increased (statistically significantly) from 28.5% in 2007 to 36.6% in 2011 to 47.8% in 2014. Thus, the proportion of the population in deprived categories has gone down over the same period. Changes within the deprived categories are however mixed. 15 The JMP document titled WASH Post-2015: Proposed indicators for drinking water, sanitation and hygiene was accessed in November 2019 at https://www.who.int/water sanitation health/monitoring/ coverage/wash-post-2015-rev.pdf. 28 Although the estimated population shares in the two most deprived categories (‘open defe- cation’ and ‘unimproved’) have decreased (statistically significantly) systematically between 2007 and 2014, the population shares in the other two deprivation categories have not. Table 4: Change in sanitation deprivation by ordinal poverty measures in Bangladesh and its divisions H PI PM PG 2007 2011 2014 2007 2011 2014 2007 2011 2014 2007 2011 2014 Barisal 66.1 60.5 46.8 47.7 44.0 32.6 35.0 32.6 23.4 25.2 23.5 16.7 (2.7) (2.1) (3.3) (1.9) (1.7) (2.9) (1.4) (1.4) (2.4) (1.1) (1.1) (1.8) Chittagong 67.1 59.2 44.9 47.0 38.9 29.2 34.8 27.3 20.2 26.2 19.6 14.9 (2.9) (2.1) (3.1) (2.8) (1.7) (2.9) (2.7) (1.5) (2.7) (2.6) (1.2) (2.5) Dhaka 74.8 68.4 56.3 50.1 42.6 33.5 36.6 29.4 21.8 27.8 21.6 15.4 (1.9) (1.9) (2.5) (1.7) (1.3) (1.8) (1.6) (1.1) (1.6) (1.4) (1.0) (1.3) Khulna 69.0 61.4 50.3 48.9 41.8 33.0 35.7 29.4 22.7 25.9 20.9 16.1 (1.8) (1.6) (2.3) (1.4) (1.2) (1.7) (1.1) (1.0) (1.4) (1.0) (0.7) (1.0) Rajshahi 74.0 61.2 52.0 55.2 41.6 34.7 43.0 29.6 24.2 34.1 21.6 17.6 (1.9) (2.3) (2.3) (1.8) (1.9) (1.8) (1.9) (1.6) (1.5) (1.9) (1.3) (1.3) Sylhet 69.9 64.6 55.0 50.1 47.1 37.9 36.8 36.2 27.9 26.5 28.9 21.9 (2.3) (1.9) (1.7) (1.9) (1.6) (1.7) (1.5) (1.4) (1.6) (1.1) (1.4) (1.4) Bangladesh 71.5 63.4 52.2 50.4 42.3 33.5 37.5 30.1 23.1 28.5 22.2 16.8 (1.0) (0.9) (1.1) (0.9) (0.7) (0.9) (0.9) (0.6) (0.8) (0.8) (0.5) (0.7) Sources: Authors’ own computations. Standard errors are reported in parentheses. Has this estimated reduction pattern been replicated within all divisions? Table 3 also presents the changes in the discrete probability distributions of three divisions: Dhaka, Rajshahi, and Sylhet.16 The estimated population shares in the ‘improved’ category have increased (statistically significantly) gradually in all three regions (Table 3), and so the shares 16 A new division called Rangpur was formed in 2010, which was a part of the Rajshahi Division. The new division did not exist during the 2007 DHS, so we had to combine this division with the Rajshahi division in the 2011 and 2014 DHS in order to preserve comparability over time. Likewise, our selected surveys were not affected by the carving out of northern Dhaka to form the Mymesingh division in 2015. 29 of deprived population have gone down. We need, however, to point out two crucial aspects. First, let us compare the reduction patterns in Dhaka and Rajshahi. The population share in the ‘improved’ category is higher in Rajshahi in 2011 and 2014 and statistically indistinguish- able in 2007, implying that sanitation deprivation in Dhaka is never lower than sanitation deprivation in Rajshahi. However, the estimated population shares in the two most deprived categories (‘open defecation’ and ‘unimproved’) are higher in Rajshahi than in Dhaka in 2007 and 2014 and statistically indistinguishable in 2011. Second, like Dhaka and Rajshahi in Ta- ble 3, sanitation deprivation in Sylhet has also improved gradually. However, the estimated population share in the poorest category (‘open defecation’) is significantly higher in 2011 and in 2014 than in 2007. A simple headcount measure, which only captures the proportion of the overall deprived population, would always overlook these substantial differences. Table 4 presents four different poverty measures for Bangladesh and for its six divisions (as per the pre-2010 administrative map). We assume the poverty threshold category to be ‘basic unsafe’. The first poverty measure is the headcount ratio (H ), which, in this context, is the population share experiencing any one of the four deprivation categories. The second measure is PI , such that PI ∈ P \ {Pα } for α ≥ 1, and is defined by the ordering weights ω I = (1, 0.75, 0.5, 0.25, 0). The third measure is PM ∈ P1 \ {Pα } for α ≥ 2 with ordering weights ω M = (1, 0.752 , 0.52 , 0.252 , 0), i.e. respecting the restrictions in Corollary 3.1, but not respecting, for instance, the restrictions in Proposition 3.1 or the restrictions in Theorem 3.1 for α ≥ 2; whereas, the fourth measure is PG ∈ PG with ordering weights ω G = (1, 0.4, 0.15, 0.05, 0), i.e. respecting the restrictions in Proposition 3.1. Note that measures PM and PG give precedence to those who are in the poorer categories. All four 30 measures lie between zero and one, but we have multiplied them by one hundred so that they lie between zero (lowest deprivation) and 100 (highest deprivation). Comparisons of these measures provide useful insights, especially into the two crucial aspects that we have presented in Table 3. The headcount ratio estimate in Dhaka is statistically indistinguishable from the headcount ratio estimate in Rajshahi for 2007, despite deprivation in the two poorest categories being higher in Rajshahi. However, this crucial aspect is captured by the latter three measures, which show statistically significantly higher poverty estimates in Rajshahi than in Dhaka. Similarly, the headcount ratio estimate is higher in Dhaka than in Rajshahi for 2011, but the difference vanishes when poverty is assessed by the other three ordinal measures. Meanwhile, Table 4 also shows that, without accounting for the depth of deprivations, the headcount ratio ranks Dhaka as poorer than Sylhet in 2011. However, this ranking reverses as soon as we use deprivation-sensitive measures from the class P . Since most of the information in Table 4 points to experiences of poverty reduction, we further conduct pair-wise dominance tests in order to verify whether the comparisons over time are robust to all measures in class P , and in sub-classes P1 and PG . Note, in this case, that S = 5 and k = 4. The dominance test for P is based on Theorem 4.1, the dominance test for P1 is based on Corollary 4.2 (i.e. α = 1), and the dominance test for PG is based on Corollary 4.3 (i.e. α = k − 1 = 3 ). Checking pair-wise dominance for P1 requires comparing poverty levels at the following four extreme points: (1, 0, 0, 0, 0), (1, 1 2 , 0, 0, 0), (1, 2 3 2 1 , 1 , 0, 0) and (1, 4 3 3 , 4 , 4 , 0). Meanwhile, checking pair-wise dominance for PG requires comparing poverty levels at the following four extreme points: (1, 0, 0, 0, 0), (1, 1 2 , 0, 0, 0), 31 (1, 1 , 1 , 0, 0) and (1, 1 2 4 , 1 , 1 , 0). 2 4 8 Table 5: Pair-wise dominance tests for changes in poverty over time in Bangladesh and its divisions 2007–2011 2011–2014 P P1 PG P P1 PG Barisal No No No Yes Yes Yes Chittagong Yes** Yes** Yes** No No No Dhaka Yes** Yes** Yes** Yes** Yes** Yes** Khulna Yes* Yes* Yes* Yes Yes Yes Rajshahi Yes*** Yes*** Yes*** Yes Yes Yes Sylhet No No No Yes* Yes* Yes* Bangladesh Yes*** Yes*** Yes*** Yes* Yes* Yes* Sources: Authors’ own computations. A “Yes” implies reduction of the poverty levels within each region over time evaluated at the four extreme weights; a “No” implies otherwise. The levels of statistical significance for the dominance tests are: *** for 1%, ** for 5% and * for 10%. Table 5 presents the pair-wise dominance tests over time for Bangladesh and its divisions. Each cell of the table reports whether a reduction or increase in poverty within each region over time in Table 4 is robust (or not) to alternative ordering weighting vectors within the class (i.e., P ) and subclasses (i.e., P1 and P2 ) to which the ordering weighting vectors used in Table 4 belong.17 For example, it is evident from Table 4 that, according to PM , poverty in Bangladesh has fallen from 37.5 points in 2007 to 30.1 points in 2011. So we could ask: is this reduction between 2007 and 2011 robust to all P ∈ P1 ? The corresponding dominance test reported in Table 5 shows that the reduction is certainly robust with 1% significance 17 We implemented a standard intersection-union test (IUT) whose alternative hypothesis is that poverty a-vis levels at the four extreme points are jointly lower in year-region “A” (e.g. Bangladesh in 2011) vis-` “B” (e.g. Bangladesh in 2007). We reject the null (i.e. at least one poverty level in “A” is equal or higher than in “B”) in favour of this alternative only if every poverty level is lower in “A” than in “B” with a given significance level α. As explained by Berger (1997, p. 226), the IUT’s overall significance level is also α and no correction for multiple comparisons (e.g. Bonferroni, etc.) is required. 32 level. If a cell in Table 5 reports either “No” or “Yes”, but without appropriate statistical significance level, then we conclude that the relevant comparison in Table 4 is not robust. For instance, poverty reduction in Dhaka has been robust for the whole class P at the 5% level of significance throughout both time periods (2007-2011 and 2011-2014); whereas, the poverty reduction for neither Barisal nor Chittagong for the period 2011-2014 reported in Table 4 can be claimed to be robust for any of the three (sub)classes of indices considered. 6. Concluding remarks There is little doubt that poverty is a multidimensional concept and the current global devel- opment agenda correctly seeks to ‘reduce poverty in all its dimensions’. To meet this target, it is indeed important to assess poverty from a multidimensional perspective. However, one should not discount the potential interest in evaluating the impact of a targeted program in reducing deprivation in a single dimension such as educational or health outcomes and access to public services, which may often be assessed by an ordinal variable with multiple ordered deprivation categories. The frequently used headcount ratio, in this case, is ineffec- tive as it overlooks the depth of deprivations, i.e. any changes within the ordered deprivation categories. Our paper has thus posed the question: ‘How should we assess poverty when variables are ordinal?’ Implicitly, the companion question is ‘Can we meaningfully assess poverty beyond the headcount ratio when we have an ordinal variable?’ Drawing on six reasonable axiomatic properties, our answer is: ‘Poverty can be measured with ordinal variables through weighted averages of the discrete probabilities corresponding to the ordered categories.’ We refer to 33 these weights as ordering weights, which need to satisfy a specific set of restrictions in order to ensure the social poverty indices fulfil these key properties. Our axiomatically characterised class of social poverty indices has certain desirable features, such as additive decomposability and being bounded between zero (when none experiences any deprivation) and one (when everyone experiences the most deprived category). In contrast to previous attempts in the literature on poverty measurement with ordinal variables, we have gone fruitfully further in the direction of operationalising different concepts of ‘precedence to the poorer people among the poor’, which ensures that the policymaker has an incentive to assist the poorer over the less poor. We have shown that it is possible to devise reasonable poverty measures that prioritise welfare improvements among the most deprived when variables are ordinal. We have axiomatically characterised a set of subclasses of ordinal poverty measures based on different notions of precedence to those poorer among the poor. Each subclass is defined by an additional restriction on the admissible ordering weights. The precedence-sensitive measures have proven useful in the illustration pertaining to sanitation deprivation in Bangladesh by highlighting those provinces where the overall headcount improvement did not come about through reductions in the proportion of the population in the most deprived categories. Since several poverty measures are admissible within each characterised class and subclass, we have also developed stochastic dominance conditions for each subclass of poverty measures. Their fulfilment guarantees that all measures within a given class (or subclass) rank the same pair of distributions robustly. While some of these conditions represent the ordinal-variable analog of existing conditions for continuous variables in the poverty dominance literature 34 (Foster and Shorrocks, 1988b); others are, to the best of our knowledge, themselves a novel methodological contribution to the literature on stochastic dominance with ordinal variables. There has been a recent surge in the literature on multidimensional poverty measurement, especially within the counting framework. In this framework, however, it is still a usual practice to dichotomise deprivations within each dimension when using existing counting measures, ignoring the depth of deprivation across ordered categories. Future research could focus on the development of counting measures that incorporate both the depth of depriva- tions within dimensions and notions of precedence to the poorest among the poor. References Abul Naga, R. and T. Yalcin (2008). Inequality measurement for ordered response health data. Journal of Health Economics 27, 1614–25. Alkire, S. and J. E. Foster (2011). Counting and multidimensional poverty measurement. Journal of Public Economics 95 (7-8), 476–87. Allison, R. A. and J. E. Foster (2004). Measuring health inequality using qualitative data. Journal of Health Economics 23, 505–24. Apouey, B. (2007). Measuring health polarisation with self-assessed health data. Health Economics 16, 875–94. Atkinson, A. B. (1987). On the measurement of poverty. Econometrica 55 (4), 749–764. 35 Bennett, C. and C. Hatzimasoura (2011). Poverty measurement with ordinal data. IIEP- WP-2011-14, George Washington University. Berger, R. (1997). Likelihood ratio tests and intersection-union tests. In S. Panchapakesan and N. Balakrishnan (Eds.), Advances in Statistical Decision Theory and Applications, pp. 225–237. Boston: Birkhauser. Bosmans, K., L. Lauwers, and E. Ooghe (2017). Prioritarian poverty comparisons with cardinal and ordinal attributes. Scandinavian Journal of Economics 120 (3), 925–942. DOI: 10.1111/sjoe.12238. Bossert, W., S. Chakravarty, and C. D’Ambrosio (2013). Multidimensional poverty and material deprivation with discrete data. Review of Income and Wealth 59 (1), 29–43. Chakravarty, S. R. (1983). Ethically flexible measures of poverty. Canadian Journal of Economics 16 (1), 74–85. Clark, S., R. Hemming, and D. Ulph (1981). On indices for the measurement of poverty. The Economic Journal 91 (362), 515–526. Cowell, F., M. Kobus, and R. Kurek (2017). Welfare and inequality comparisons for uni- and multi-dimensional distributions of ordinal data. Mimeograph. Dhongde, S., Y. Li, P. Pattanaik, and Y. Xu (2016). Binary data, hierarchy of attributes, and multidimensional deprivation. Journal of Economic Inequality 14, 363–378. Ebert, U. (2007). Ethical inequality measures and the redistribution of income when needs differ. Journal of Economic Inequality 5, 263–278. 36 Fields, G. (2001). Distribution and development: a new look at the developing world. MIT Press. Fishburn, P. and I. Lavalle (1995). Stochastic dominance on unidimensional grids. Mathe- matics of Operations Research 20 (3), 513–525. Fleurbaey, M. (2015). Equality versus priority: how relevant is the distinction? Economics and Philosophy 31 (2), 203–217. Foster, J., M. McGillivray, and S. Seth (2012). Rank robustness of composite indices: Dom- inance and ambiguity. OPHI Working Paper 26b, University of Oxford. Foster, J. E., J. Greer, and E. Thorbecke (1984). A class of decomposable poverty measures. Econometrica 52 (3), 761–6. Foster, J. E. and A. K. Sen (1997). On Economic Inequality. Oxford University Press. Foster, J. E., S. Seth, M. Lokshin, and Z. Sajaia (2013). A unified approach to measuring poverty and inequality: Theory and practice. Washington DC: The World Bank. Foster, J. E. and A. F. Shorrocks (1988a). Poverty orderings. Econometrica 56 (1), 173–7. Foster, J. E. and A. F. Shorrocks (1988b). Poverty orderings and welfare dominance. Social Choice and Welfare 5, 179–98. Gravel, N., B. Magdalou, and P. Moyes (2015). Ranking distributions of an ordinal attribute. Technical report, halshs-01082996v2. Gravel, N., B. Magdalou, and P. Moyes (2018). Inequality measurement with an ordinal and continuous variable. https://doi.org/10.1007/s00355-018-1159-8. 37 Guenther, R. and J. Lee (1988). Partial Differential Equations of Mathematical Physics and Integral Equations. Dover Publications. Hammond, P. J. (1976). Equity, Arrow’s conditions, and Rawls’ difference principle. Econo- metrica 44 (4), 793–804. Kobus, M. (2015). Polarisation measurement for ordinal data. Journal of Economic Inequal- ity 13 (2), 275–97. Kobus, M. and P. Milos (2012). Inequality decomposition by population subgroups for ordinal data. Journal of Health Economics 31, 15–21. Lazar, A. and J. Silber (2013). On the cardinal measurement of health inequality when only ordinal information is available on individual health status. Health Economics 22 (1), 106–13. Levy, H. (2006). Stochastic Dominance: Investment decision making under uncertainty (2nd ed.). Springer. Mendelson, H. (1987). Quantile- preserving spread. Journal of Economic Theory 42, 334–51. NIPORT, Mitra and Associates, and ICF International (2009). Bangladesh demographic and health survey 2007. Report, Dhaka, Bangladesh, and Calverton, Maryland, USA. NIPORT, Mitra and Associates, and ICF International (2013). Bangladesh demographic and health survey 2011. Report, Dhaka, Bangladesh, and Rockville, Maryland, USA. NIPORT, Mitra and Associates, and ICF International (2016). Bangladesh demographic and health survey 2014. Report, Dhaka, Bangladesh, and Rockville, Maryland, USA. 38 Oui-Yang, K. (2018). Lexical measures of social inequality from Pigou-Dalton to Hammond. Review of Income and Wealth . DOI: 10.1111/roiw.12402. Parfit, D. (1997). Equality and priority. Ratio 10 (3), 202–221. Permanyer, I. and C. D’Ambrosio (2015). Measuring social polarization with ordinal and categorical data. Journal of Public Economic Theory 17 (3), 311–27. Ravallion, M. (1994). Poverty Comparisons. Switzerland: Harwood Academic Press. Roberts, F. S. (1979). Measurement Theory with Applications to Decisionmaking, Utility, and the Social Sciences, Volume 7 of Encyclopedia of Mathematics and its Applications. Reading, Mass.: Addison-Wesley. Sen, A. K. (1976). Poverty: An ordinal approach to measurement. Econometrica 44 (2), 219–231. Seth, S. and M. McGillivray (2018). Composite indices, alternative weights, and comparison robustness. Social Choice and Welfare 51 (4), 657–679. Seth, S. and G. Yalonetzky (2018). Assessing deprivation with ordinal variables: Depth sensitivity and poverty aversion. OPHI Working Paper 123. University of Oxford. Shorrocks, A. F. (1995). Revisiting the sen poverty index. Econometrica 63 (5), 1225–1230. Stevens, S. S. (1946). On the theory of scales of measurement. Science, New Se- ries 103 (2684), 677–680. Thon, D. (1979). On measuring poverty. Review of Income and Wealth 25 (4), 429–439. 39 World Bank (2017). Monitoring global povety: Report of the commission on global poverty. Washington, DC: World Bank. Yalonetzky, G. (2012). Poverty measurement with ordinal variables: A generalization of a recent contribution. ECINEQ WP 2012–246. Yalonetzky, G. (2013). Stochastic dominance with ordinal variables: Conditions and a test. Econometric Reviews 32 (1), 126–63. Zheng, B. (1997). Aggregate poverty measures. Journal of Economic Surveys 11 (2), 123– 162. Zheng, B. (2011). A new approach to measure socioeconomic inequality in health. Journal of Economic Inequality 9 (4), 555–77. Appendices Appendix A1. Proof of Theorem 2.1 It is straightforward to check that each poverty measure in Equation 2.1 satisfies the four properties: OMN, SCD, FOC and SUD. We now prove the necessary part, namely, if a poverty measure satisfies these four properties, then it takes the functional form in Equation 2.1. First, we define S ∈ N \ {1} basis vectors: es = (es s s t 1 , . . . , eS ) ∈ P ∀s = 1, . . . , S , such that es = 1 ∀s and es = 0 ∀s = t. For a given ck ∈ C−S , denote P (es , ck ) = ωs for some ωs ∈ R+ for every s = 1, . . . , S . 40 Next, suppose a society with probability distribution p = (p1 , . . . , pS ) ∈ P is divided into M = S mutually exclusive and collectively exhaustive sub-groups, such that the probability distribution of each subgroup is ps = (ps s s 1 , . . . , pS ) = e for every s = 1, . . . , S (= M ). As each subgroup contains the population in each category, naturally, the population share of each subgroup is πs = ps for every s = 1, . . . , S (= M ). Then, by SUD: S S S s s P (p, ck ) = πs P (p , ck ) = ps P (e , ck ) = ps ωs . (A1) s=1 s=1 s=1 Now, consider any two p, q ∈ P, such that ps = 1 for some s ≤ k and qs = 1 for some s > s. Naturally, pt = qt = 0 for all t = {s, s }. Clearly, by OMN, we have P (q, ck ) < P (p, ck ). Combined with Equation A1, we then obtain: ωs < ωs . (A2) The relationship in Equation A2 holds for any s, such that s ≤ k and s < s . In other words, ωs−1 > ωs > ωs for all s = 2, ..., k and even for any s > k , whenever k ≥ 2. When S = 2, then k = 1 and so ω1 > ω2 . We next use property SCD. Suppose, k = 1. Then, property SCD leads to P (p, ck ) = p1 and Equation A1 yields S P (p, c1 ) = ps ωs = p1 . (A3) s=1 Note, by definition, that 0 ≤ p1 ≤ 1 and so 0 ≤ P (p, ck ) ≤ 1. Consider some p ∈ P, such that p1 = 1 and ps = 0 for all s = 1. Clearly, from Equation A3, ω1 = 1. Moreover, from 41 Equation A2, it follows that 1 > ωs > ωs for all s = 2, . . . , k and for any s > k whenever k ≥ 2. In order to complete the proof, we need to show that ωs = 0 for all s > k . For this purpose, consider some p, q ∈ P, such that pt = qu = 1 for some t > u > k . Certainly, ps = qs = 0 ∀s ≤ k . By property FOC, we then require P (p, ck ) = P (q, ck ). Thus, from Equation A1, we obtain P (p, ck ) = ωt = ωu = P (q, ck ) for t > u > k . Since, ps = qs = 0 ∀s ≤ k , it follows that p1 = q1 = 0. Consider k = 1. Then, by property SCD, we must have P (p, ck ) = P (q, ck ) = 0. Hence, it must be the case that ωs = 0 for all s > k , which completes our proof. Appendix A2. Proof of Theorem 3.1 The sufficiency part is straightforward. We prove the necessity part as follows. Suppose k ≥ 2 and α ∈ N such that 1 ≤ α ≤ k − 1. Now, suppose, p and q are obtained from p ∈ P as follows. Consider some s < t ≤ k < S . Now, p is obtained from p, such that ps = ps − while pu = pu ∀u = {s , s + 1}. Naturally, ps +1 = ps +1 + . Similarly, q is obtained from p, such that qt = pt − while qu = pu ∀u = {t, t } for some t = min{t + α, S }. Naturally, again, qt = pt + . By property PRE-α, we know that P (p , ck ) < P (q , ck ). (A1) 42 Combining Equation 2.1 and Equation A1, we get ωs +1 − ωs − ωt + ωt < 0. Substituting t = s + 1 = s for any s = 2, . . . , k , we obtain ωs−1 − ωs > ωs − ωt . (A2) First, suppose t = s + α ≤ k < S or s ≤ k − α. Then, ωt = ωs+α > 0 by Theorem 2.1 and Equation A2 can be expressed as ωs−1 − ωs > ωs − ωs+α for all s = 2, . . . , k − α. Second, suppose t = min{s + α, S } > k or s > k − α. We know that ωs = 0 for all s > k by Theorem 2.1 and so Equation A2 can be expressed as ωs−1 − ωs > ωs or ωs−1 > 2ωs for all s = k − α + 1, . . . , k . This completes the proof. Appendix A3. Proof of Proposition 3.1 Let us first prove the sufficiency part. Suppose k ≥ 2. We already know from Theorem 2.1 that ωs−1 > ωs > 0 ∀s = 2, . . . , k and ωs = 0 ∀s > k . Suppose additionally that ωs−1 > 2ωs ∀s = 2, . . . , k . Alternatively, ωs−1 − ωs > ωs ∀s = 2, . . . , k . For any p, p , q ∈ P, for any ck ∈ C−S and for some ∈ (0, 1), let k ≥ t ≥ v + α and suppose p is obtained from p, such that pv = pv − and pv+α = pv+α + , while pu = pu ∀u = {v, v +α}; and q is obtained from p, such that qt = pt − and qt+β = pt+β + for some β ∈ N, while qu = pu ∀u = {t, t + β }. With the help of Equation 2.1, we get: P (p , ck ) − P (q , ck )] = [ωv+α − ωv − ωt+β + ωt ] = [(ωt − ωt+β ) − (ωv − ωv+α )]. (A1) 43 By assumption of the sufficiency part: ωs−1 − ωs > ωs ∀s = 2, . . . , k . Combining this assumption with the weight restrictions from Theorem 2.1 we can easily deduce that (ωv − ωv+α ) > (ωv − ωv+1 ) > ωv+1 > ωv+α . Hence, (ωv − ωv+α ) > ωv+α . Since v + α ≤ t ≤ k and ωs−1 > ωs > 0 ∀s = 2, . . . , k , it also follows that ωv+α ≥ (ωt − ωt+β ). Hence, (ωv − ωv+α ) > (ωt − ωt+β ) and P (p , ck ) < P (q , ck ). Next we prove the necessity part starting with Equation A1. By property PRE-G, we know that P (p , ck ) < P (q , ck ). Thus, ωv − ωv+α > ωt − ωt+β . (A2) Now the inequality in Equation A2 must hold for any situation in which t ≥ v + α, including the comparison of the minimum possible improvement for the poorer person, given by ωv − ωv+1 (i.e. with α = 1), against the maximum possible improvement for the less poor person, given by ωt − ωt+β with t = v + 1 and t + β > k . Inserting these values into Equation A2, bearing in mind that ωt+β = 0 when t + β > k , yields ωv − ωv+1 > ωv+1 . Substituting v = s − 1 for any s = 2, . . . , k yields ωs−1 − ωs > ωs . Hence, ωs−1 > 2ωs for all s = 2, . . . , k . Appendix A4. Proof of Theorem 4.1 We denote the difference operator by ∆ and use the following handy definitions: For some s p, p ∈ P, define ∆Pk ≡ P (p, ck ) − P (p , ck ), ∆ps ≡ ps − ps , and ∆Fs ≡ =1 ∆p . 44 We first prove the sufficiency part. From Theorem 2.1, we know that ωs = 0 for all s > k . k Thus, Equation 2.1 may be presented using the difference operator as ∆Pk = s=1 ωs ∆ps . Using summation by parts, also known as Abel’s lemma (Guenther and Lee, 1988) or formula (Fishburn and Lavalle, 1995, p. 518), it follows that k −1 ∆Pk = [ωs − ωs+1 ]∆Fs + ∆Fk ωk . (A1) s=1 We already know from Theorem 2.1 that ωk > 0 and ωs − ωs+1 > 0 ∀s = 1, . . . , k − 1. Therefore, clearly from Equation A1, the condition that ∆Fs ≤ 0 ∀s ≤ k and ∆Fs < 0 for at least one s ≤ k is sufficient to ensure that ∆Pk < 0 ∀P ∈ P and for a given ck ∈ C−S . We next prove the necessity part by contradiction. Consider the situation, where ∆Ft > 0 for some t ≤ k , ∆Fs ≤ 0 for all s ≤ k but s = t, and ∆Fs < 0 for some s ≤ k but s = t. For a sufficiently large value of ωt − ωt+1 in Equation A1, it may always be possible to find that ∆Pk > 0. Or, consider the situation ∆Fs = 0 for all s ≤ k . In this case, ∆Pk = 0. Hence, the necessary condition requires both ∆Fs ≤ 0 for all s ≤ k and ∆Fs < 0 for some s ≤ k . This completes the proof. Appendix A5. Proof of Theorem 4.2 Consider some k ≥ 2 and some α ∈ [1, k − 1] ⊆ N. Let Ωk α be the set of all S -dimensional 45 ordinal weighting vectors corresponding to subclass Pα (see Theorems 3.1 and 2.1), so that:        (ω1 , . . . , ωS ) | ω1 = 1; ωs−1 > 2ωs − ωs+α ∀s = 2, . . . , k − α; ωs−1 > 2ωs > 0      ∀s = k − α + 1, . . . , k ; and ωs = 0 ∀s > k whenever α ≤ k − 2    k Ωα = .   (ω1 , . . . , ωS ) | ω1 = 1; ωs−1 > 2ωs > 0 ∀s = 2, . . . , k ; and ωs = 0 ∀s > k             whenever α = k − 1 ¯ k as: Now define Ω α        (ω1 , . . . , ωS ) | ω1 = 1; ωs−1 ≥ 2ωs − ωs+α ∀s = 2, . . . , k − α; ωs−1 ≥ 2ωs ≥ 0      ∀s = k − α + 1, . . . , k ; and ωs = 0 ∀s > k for α ≤ k − 2    ¯ k Ωα = .   (ω1 , . . . , ωS ) | ω1 = 1; ωs−1 ≥ 2ωs ≥ 0 ∀s = 2, . . . , k ; and ωs = 0 ∀s > k             for α = k − 1 ¯ k is bounded by a set of linear constraints and it is a convex For given values of k and α, Ω α ¯ k . For ¯ 1, . . . , ω hull of k extreme points or extreme ordering weighting vector, denoted by ω ¯ r , where r ∈ {1, 2, . . . , k }, maximum feasible positive weights are the rth extreme point ω assigned to the first r elements and minimum feasible weights are assigned to the rest of ¯k the elements (where feasibility refers to satisfying the linear constraints defining the set Ωα ), r r such that ωs = ω ¯s > 0 for all s = 1, . . . , r and ωs = ω ¯s = 0 for all s = r + 1, . . . , S .18 Any ¯ k is a convex combination of these k extreme ordering weighting vectors, such that ω ∈Ω α k k ω= ¯ r, r=1 θr ω where r=1 θr = 1 and θr ≥ 0 for all r = 1, . . . , k . The restrictions on the aforementioned θr ’s may be verified as follows. Since, ω1 = 1 and r k r ¯1 ω = 1 for all r by Theorem 2.1, then plugging these values in ω1 = ¯1 r=1 θr ω yields 18 The proof is similar in spirit to the proof of Proposition 1 in Seth and McGillivray (2018). 46 k r=1 θr = 1. Next, define:   ωs − 2ωs+1 + ωs+α+1 ∀s = 1, . . . , k − α − 1 and ωs − 2ωs+1         us = ∀s = k − α, . . . , S − 1 whenever α ≤ k − 2 .        ω − 2ω  s ∀s = 1, . . . , S − 1 whenever α = k − 1 s+1 k r r Since, ωs = ¯s r=1 θr ω ¯s for all s and ω = 0 for all s > r, we obtain:   k k k r r r ¯s r=s θr ω −2 ¯s r=s+1 θr ω + r=s+1+α θr ω ¯s when s ≤ k − α − 1    +1 +1+α us = .  k k  r r r=s θr ω ¯s −2 ¯s r=s+1 θr ω otherwise   +1 Rearranging the right hand side of us yields: s+ α k s r r r r r us = ¯s θs ω + ωs θr [¯ − ωs 2¯ +1 ] + θr [¯ ωs − 2¯ ωs ¯s +1 + ω +1+α ], (A1) r=s+1 r=s+1+α where the third component on the right hand side of Equation A1 exists only when s + α +1 ≤ ¯ k , the second k . Given that the extreme weighting vectors satisfy the equality restrictions in Ω α component and the third component (whenever in existence) on the right-hand side must be s ¯s equal to zero. Hence, us = θs ω for all s, which is unrelated to the restrictions of Theorem ¯ k and also ω 3.1. Given that us ≥ 0 for all s = 1, . . . , k by the definition of Ω ¯ss > 0 (i.e., α s = r) for all s = 1, . . . , k by the definition of extreme weighting vectors, then, substituting r = s, it must be the case that θr ≥ 0 for all r = 1, . . . , k . The primary difference between Ωk ¯k ¯k α and Ωα is that the elements of Ωα satisfy the non-strict versions of the inequality restrictions prescribed by Theorem 3.1. Clearly, Ωk ¯k α ⊂ Ωα and k therefore any ω ∈ Ωk ¯ k , i.e., ω = ¯ 1, . . . , ω α is also a convex combination of ω ¯ r, r=1 θr ω such k that r=1 θr = 1. However, the strict inequality constraints in Ωk α require that us > 0 for all s = 1, . . . , k , which implies that θr > 0 for all r = 1, . . . , k for Ωk α. 47 For any p ∈ P, let us denote the poverty level evaluated at some ω as P (p, ck ; ω ) and, given k k additive decomposability, we must have P (p, ck ; ω ) = ¯ r ), r=1 θr P (p, ck ; ω where r=1 θr = ¯ r) ¯ r ) ≤ P (q, ck ; ω 1 and θr > 0 for all r = 1, . . . , k . Clearly, for some p, q ∈ P, if P (p, ck ; ω for all r = 1, . . . , k , with one strict inequality, then it follows directly that P (p, ck ; ω ) < P (q, ck ; ω ) for any ω ∈ Ωk α or, equivalently, for all P ∈ Pα for a given ck . This proves sufficiency. ¯ r ) for some r ∈ {1, 2, . . . , k }. ¯ r ) > P (q, ck ; ω In order to prove necessity, suppose P (p, ck ; ω Then, for a sufficiently large value of θr (i.e., for θr → 1), it is always possible to have ¯ r ) for all r = ¯ r ) = P (q, ck ; ω P (p, ck ; ω ) > P (q, ck ; ω ). Furthermore, in case P (p, ck ; ω 1, . . . , k , then certainly P (p, ck ; ω ) = P (q, ck ; ω ) for all ω ∈ Ωk α . Hence, the conditions, ¯ r ) for all r = 1, . . . , k , with one strict inequality, are also necessary. ¯ r ) ≤ P (q, ck ; ω P (p, ck ; ω Finally, the solutions for the k extreme points are obtained as follows. By Theorem 2.1, r r ¯1 ω ¯s = 1 for all r = 1, . . . , k . From above, we already know that ω = 0 for all s = r + 1, . . . , S r and for every r. Therefore, for every r ∈ {1, . . . , k }, ω ¯s = 0 for all s > r. We also know that r ω ¯s > 0 for all s ≤ r. Now, for any α and for every r ∈ {2, . . . , k }, we obtain the following r r r r r r r r r system of (r − 1) equations: ω ¯1 ω2 = 2¯ −ω ¯ 2+α, ω ¯2 = 2¯ ω3 −ω ¯ 3+α, . . . , ω ¯r −α−1 = 2¯ ωr −α − ω ¯r , r r r r r r r r ω ¯r −α = 2¯ ωr −α+1 , . . . , ω ¯r ωr −2 = 2¯ −1 , ω ¯r −1 = 2¯ ωr . There are (r − 1) unknowns: ω ¯2 ¯r ,...,ω , r r r ¯1 since ω = 1. Solving the system of equations, we obtain ωs ¯s =ω for all s = 2, . . . , r and for all r = 1, . . . , k in Equation 4.1. This completes the proof of the theorem. 48 Appendix A6. Proof of Corollary 4.2 We use the difference operators defined in Appendix A4. Summing by parts the first com- ponent on the right-hand side of Equation A1 and then rearranging the terms, we obtain: k −1 s k ∆Pk = {[ωs − ωs+1 ] − [ωs+1 − ωs+2 ]} ∆F + ωk ∆Fs . (A1) s=1 =1 s=1 We know from Corollary 3.1 that [ωs − ωs+1 ] − [ωs+1 − ωs+2 ] > 0 ∀s = 1, ..., k − 1. Likewise, s by Theorem 2.1, ωk > 0. Therefore =1 ∆F ≤ 0 for all s = 1, ..., k with at least one strict inequality is sufficient to ensure that, for any p, q ∈ P, P (p, ck ) < P (q, ck ) ∀P ∈ P1 . On the s contrary, suppose =1 ∆F > 0 for some s ∈ {1, . . . , k }. Now, note that there is no further restriction on whether any of the weight functions in Equation A1 is strictly greater than s the others. Then, attaching a sufficiently large weight to this component (i.e. =1 ∆F > 0 for some s ∈ {1, . . . , k }) may result in P (p, ck ) > P (q, ck ) for some P ∈ P1 . Hence, the conditions are also jointly necessary, which completes the proof. 49