Syndicate content

Global poverty today, the 1908 winter in St. Petersburg, and ‘controversy bias’

Francisco Ferreira's picture

Robert Allen’s recent AER paper on “Absolute Poverty: When Necessity Displaces Desire” is a fascinating read, on many levels. The paper uses linear programming (LP) to compute (four variants of) least-cost diets for twenty countries, using prices from the International Comparisons Project (ICP) microdata. To the resulting least-cost food budgets, estimates of non-food costs covering housing, fuel, lighting, clothing and soap are added, generating “basic need poverty lines” (BNPL) for each country.

The paper makes many interesting claims. One is that, although linear programming algorithms generate unrealistic diets in rich countries, among very poor people in the world’s poorest countries deprivation is so constraining that little latitude is left to taste, and the LP predictions turn out to be reasonable approximations of actual diets. This is an interesting hypothesis, although the actual diets to which the LP predictions are compared are averages for 1961, from the FAO Food Balances Sheets. As Martin Ravallion notes in this Comment, it should have been possible to use data for - or near - 2011 and, furthermore, what appears as a reasonable match to Allen can seem like large differences to others.

In this post, however, I want to focus on a different claim (which Martin also addresses), namely that “[t]his approach is superior to the World Bank’s $1-a-day line…” (Allen’s abstract, p.3690). Bob Allen is a distinguished economic historian, with considerable influence and a well-deserved reputation in the profession. He was a member of the Bank’s own Commission on Global Poverty (as was I, for full disclosure). His views clearly deserve to be taken seriously. Yet, in claiming that his BNPL approach is strictly superior to the “Bank’s approach” - originally due to Ravallion et al. (1991) – the paper fatally over-reaches.  This is because the results depend crucially on a number of assumptions that are at least as arbitrary and problematic as any used in the “Bank’s method”. Consider these three examples:

  1. Rather than relying on countries’ own assessments of what defines poverty in their contexts (the national poverty lines used by the Bank’s approach), the BNPL relies on costing various food and non-food items. These include the cost of renting 3 square meters of shelter (as opposed to, say, 2.5 or 3.5, but never mind…). For six countries, the ICP reports rents for “traditional” housing. Elsewhere, we are referred to estimates on an online appendix, arrived at with the aim of “finding market (not subsidized) rents in the poorest districts of the capital or other large cities.” (p.3710) Yet most extremely poor people live in rural areas, where ‘rental costs’ must be considerably lower than even those in the poorest capital city slum.  How much lower would the BNPL be if one could use rental costs better aligned with where the poor live?
  2. Much is also made of the fact that, unlike the World Bank Poverty Line (WBPL), the BNPL adjusts for differences in the energy and clothing needs arising from climate differences. Sounds great.  In practice, this is done by comparing expenditures on fuel, clothing, bedding and footwear between a sample of workers in St. Petersburg in 1907-1908, with a sample of cotton mill workers in Bombay, in 1921-22. Not a sample of poor Indians today, or in 1921. It is a long time ago, so we are told that these people were all poor. In fact, the author’s identifying assumption is that “these low-income workers were at similar levels of deprivation, so that differences in their expenditures represent responses to climate, and not to real income or price differences”.  Perhaps. But should we not wonder whether the BNPL might look different had Allen had different samples at his disposal?  Say: cocoa farmers in Cote d’Ivoire in 1933 (tropical) and coal miners in Mongolia in 1927 (very cold winters)? Would such a comparison have yielded a similar climate adjustment?
  3. But there is at least one more choice that casts doubt on claims of scientific precision and methodological superiority: the set of nutritional requirements to be imposed as constraints to the LP algorithm. Allen’s published version clearly acknowledges this, and considers four options: (i) a 1,700 calorie model; (ii) a calorie-protein-fat (CPF) model; (iii) a “basic model”, and (iv) a “full course” model.  The last two differ from the CPF model by adding the recommended daily allowances (RDAs) of various minerals and vitamins. Averaging the CPF BNPLs across the fourteen countries Allen calls “developing” in his sample of 20 countries yields a daily cost of U$1.88 per day, very near the WBPL of U$1.90.Key to the paper’s claim that the BNPL does a better job than the Bank’s line, and finds more poverty, is then a rejection of the CPF diet: “…since people eating a CPF diet suffer many nutritional deficiencies, it is not a good poverty line. Instead, the line implied by the basic diet is preferable.” (p.3714) This line is $2.63 per day, naturally leading to higher poverty estimates than the $1.90 line in these countries.
    Yet, the RDAs for nutrients such as iron, folate, thiamine, niacin, vitamin C and vitamin B12 are not absolute thresholds between deprivation and prosperity, or between death and survival, or even between ill-health and good health.  Allen knows and recognizes this: “This latitude arises because the requirements for micronutrients are set at levels that the medical profession judges to be needed for healthy living. These levels exceed intake needed to prevent acute deficiency symptoms, as we have mentioned.” (pp.3708-09, emphasis added)
    In fact, this used to concern the author so much that, in the working paper version of this article, he presented five, not four, least-cost diet “models”. Between the CPF and the basic model, there was a “reduced basic model”, consisting of “CPF requirements, plus half of the Indian recommended daily allowances of iron, folate, thiamine, niacin, and the RDA of vitamins C and B12.” (p.7) This model was used because “… the Indian RDAs may be too stringent” (p.10). 
    Interestingly, the BNPL implied by this reduced basic model for non-OECD countries was exactly $1.90 per day, as reported on Table 12 (of the working paper). As the author wrote in that version: “Then we can see how the LPPL [his original name for the BNPL] relates to the World Bank poverty line. Do they equal each other? If so, with what nutritional requirements? The answer is that the same range of requirements that explains the behavior of the poor – the reduced basic and the basic models – rationalizes the $1.90 per day poverty line.” (p.14) In fact, that version of the paper concluded that “the LPPL is a new basis for the World Bank Poverty Line” (p.17)
    Now, to be fair, the mysterious disappearance of the reduced basic model was not the only change from the working paper to the published article. The climate adjustment to the costs of energy and clothing appear to be new in the published version, and housing costs are calculated differently. No one disputes that authors are entitled to rewrite a paper before it is finally published, or that we should take the published version as representing their final views. But given that, in this case, the published version reiterates that there is latitude around the RDAs for a number of vitamins and minerals, and that they exceed the lowest necessary intakes, it is puzzling that the reduced basic model – one of the two originally preferred by the author – is now absent. Whatever the actual reason, the shift certainly made it easier for this version to be more critical of the World Bank’s established method, which is in turn likely a plus for publication and ‘buzz’, in a world that values controversy more than agreement.
So, are Allen’s least-cost diets superior to the WBPL?
Thirty years ago, Tony Atkinson wrote in this Econometrica paper “that there is likely to be a diversity of judgments affecting all aspects of measuring poverty and that we should recognize this explicitly in the procedures we adopt. This will lead to less all-embracing answers. […] But such partial answers are better than no answers” (p.750). He was proposing the use of stochastic dominance techniques to make robust poverty comparisons when there are disagreements about poverty lines and indices. No matter: we would be well-advised today to be guided by the spirit of his remark. 

The point of my three examples above was to illustrate that there is an unavoidable arbitrariness in the assumptions that must inevitably be made in setting a single monetary threshold to separate the poor from the non-poor.  If you don’t like PPP exchange rates (but don’t mind using the urban-biased prices on which they are based), then you may prefer the BNPLs to the WBPL. If you don’t trust anchoring your choice of poverty line on those chosen by the countries themselves, you might also prefer the BNPL.  But you will need to defend the choice of a particular set of recommended daily allowances for things like thiamin and niacin – a subject on which no lesser an authority than Bob Allen has evidently been of two minds. You will need to choose the number of square meters a poor person should be allowed to live on, and whether pricing them using urban slums when most poverty is rural makes any sense. You will need to be willing to peg your climate adjustments to a comparison between long dead citizens of Mumbai and St. Petersburg.  You will need to acknowledge that, under equally plausible but slightly different assumptions – including some in an earlier version of this paper – the method is actually very supportive – indeed “provides a new basis for” – the World Bank Poverty Line.

Personally, I quite like the original approach of Ravallion, Datt and van de Walle (1991), also followed by various papers by Ravallion and Chen, of delegating the identification decision to the poorest countries in the world.  They know best what poverty is for them. And we then use some average of those lines as our threshold for identifying those living in extreme absolute poverty in the world.  Of course, poverty means different things in other contexts, and the $1.90 is increasingly only one element – albeit a key one – in a richer menu of approaches to quantifying global poverty. This may appear less scientifically and medically precise than some linear programming algorithm but, as Allen’s paper illustrates, such precision is so reliant on multiple questionable assumption as to be ultimately elusive.

Add new comment