When Method Matters: Monitoring Poverty in Bangladesh

Author(s): Martin Ravallion and Binayak Sen

Reviewed work(s):
Source: Economic Development and Cultural Change, Vol. 44, No. 4 (Jul., 1996), pp. 761-792
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/1154351 .
Accessed: 18/01/2013 00:14

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to
Economic Development and Cultural Change.

http://www.jstor.org

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
When Method Matters: Monitoring
Poverty in Bangladesh*

MartinRavallion
WorldBank

Binayak Sen
Bangladesh Institute of DevelopmentStudies

I. Introduction
Debates about the methods of poverty measurementare common;
views differ on how individualwelfare should be measured,how pov-
erty lines should be set, and what poverty measures should be used.
Recent controversies over Bangladesh'sprogress in reducingpoverty
illustratewell some of the main issues. There have been a numberof
claims, based on household surveys, that poverty fell in the 1980s.
However, questions have been raisedabout the comparabilityof those
surveys, the methodologiesused in analyzingthe data, and their con-
sistency with otherpoverty data such as real wages for unskilledlabor.
This article reviews the methodologicalissues and provides new
estimates of a range of poverty measuresfor urbanand ruralareas for
the period 1983-92. Methodologicalchoices in poverty measurement
will depend on the purposes of measurementand the data available.
Here we assume that the aim is to make consistent comparisons of
absolute levels of consumptionover time and space. We argue that
some past methods that have been popularin Bangladesh(and else-
where) are unlikely to do this well and can be improvedon by using
availabledata. Whileour mainaim is to throwlight on the methodolog-
ical issues, we will also look into some implicationsfor understanding
the proximatecauses of changes in poverty measures.
Past studies for Bangladeshare reviewed in Section II; we argue
that some of the methods used are preferableto others. In Section III
we give our own estimates and discuss their robustness and consis-
tency with data on real wages. Section IV examines proximatecauses
of the changes in poverty measures observed over this period and
implicationsfor future progress in reducingpoverty.
? 1996by The Universityof Chicago.All rightsreserved.
0013-0079/96/4404-0008$01.00

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
762 EconomicDevelopmentand CulturalChange

II. MeasurementControversies
Bangladeshprovides a good case study in the methodologicalissues
of poverty measurement.We shall not attempt a complete review of
past estimates for Bangladesh;rather,we focus on possible limitations
of past methodologies, to the extent that they have bearing on the
present study. The principal survey instrumentfor making poverty
comparisons over time in Bangladeshis the Household Expenditure
Survey (HES) conducted by the Bangladesh Bureau of Statistics.
There have been five such surveys since 1980, in 1981/82, 1983/84,
1985/86,1988/89,and 1991/92.

Past Estimates
Given the numerouschoices of data and methods that go into poverty
monitoring(as in almost all empiricalresearch),it would be reassuring
if different studies gave similar results. Table 1 shows five sets of
estimates of the head-countindex of poverty-the percentage of the
populationdeemed to be below the poverty line-for various years in
the 1980s.One of these is producedby the governmentof Bangladesh.1

TABLE 1
RECENTESTIMATESOF THE HEAD-COUNT INDEX OF POVERTY
IN THE1980s
FORBANGLADESH

ESTIMATEDPERCENTAGE
OF THE POPULATIONBELOWTHE POVERTYLINE

1. Rahman 2. Ahmed 4. Hossain 5. Sen

and Haque et al. 3. BBS and Sen and Islam
Urban:
1981/82 50.7 65.3 66.0 N.A. 48.4
1983/84 39.5 N.A. 66.0 N.A. 42.6
1985/86 29.1 66.8 56.0 N.A. 30.6
1988/89 N.A. N.A. 44.0 N.A. 33.4
Rural:
1981/82 79.1 71.8 73.8 65.3 N.A.
1983/84 49.8 N.A. 57.0 50.0 N.A.
1985/86 47.1 51.6 51.0 41.3 N.A.
1988/89 N.A. N.A. 48.0 43.8 N.A.

SOURCEs.-Atiq Rahmanand TrinaHaque, "Povertyand Inequalityin Bangladesh

in the Eighties:An Analysisof Some Recent Evidence," ResearchReportno. 91 (Ban-
gladesh Instituteof DevelopmentStudies, Dhaka, 1988);AkhterU. Ahmed, HaiderA.
Khan, and RajanK. Sampath,"Povertyin Bangladesh:Measurement,Decomposition
and IntertemporalComparison,"Journal of DevelopmentStudies 27 (1991): 48-63;
Bangladesh Bureau of Statistics (BBS), Report of the Household Expenditure Survey,
1988-89 (Dhaka:BBS, 1991);MahabubHossain and BinayakSen, "RuralPoverty in
Bangladesh:Trends and Determinants,"Asian DevelopmentReview 10 (1992): 1-34;
Binayak Sen and Quazi TowfiqulIslam, "MonitoringAdjustmentand Urban Poverty
in Bangladesh:Issues, Dimensions,Tendencies," in MonitoringAdjustmentand Pov-
erty in Bangladesh, Reporton the FrameworkProject, CIRDAPStudy Series no. 160
(Dhaka:Centrefor IntegratedRuralDevelopmentfor Asia and the Pacific, 1993).

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 763

Each is based on the same primarydata source-consumption expen-

ditures reportedfrom the HES for each year-and approximatelythe
same food-energyrequirement-around 2,100-2,200 calories per per-
son per day. Yet they differ considerably;across studies for 1985/86
the estimates of the proportionof the urban population that is poor
vary from 29% to 67%, and for ruralareas from 41% to 52%. These
discrepancies reflect a numberof differences in methodology, which
we discuss below. A common feature of two of the four studies of
rural poverty is that they suggest a falling poverty incidence during
the 1980s, with all four indicatinga particularlysharpfall in the early
1980s. Three studies also indicate falling poverty incidence in urban
areas duringthe early 1980s.
The comparisons between urban and rural poverty are also of
interest. As elsewhere, poverty indicators have generally suggested
that ruralpoverty is the greaterproblemin Bangladesh.2 However, two
of the studies suggest a reversalaroundthe mid-1980sin the historical
poverty orderingof the two sectors; all three indicatea higherpoverty
incidence in rural areas in 1981/82,but Akhter Ahmed et al.'s study
and that by the BangladeshBureauof Statistics (BBS) suggest a higher
incidence in urbanareas in 1985/86.
Some observers have disputed the claims that the incidence of
poverty in Bangladeshfell duringthe 1980s.3 Other-albeit partial-
indicatorsof poverty do not suggest sustainedprogress in the 1980s.
For example, the real wage rate of agriculturallabor rose in the early
1980s, peaking in about 1985, but appearsto have fallen since then,4
though there is also controversy aroundthis point, which we will try
to clear up later. Also overall economic growth-and growth in key
sectors for the poor (notably agriculture)-has been uneven in the
1980s.5Duringthe period 1981/82to 1985/86,the rate of growthin real
consumptionper capitaimpliedby the HES is 10%per year, while the
national accounts suggest a much lower rate, about 0.5% per year;6
both may be wrong-but the discrepancy is worrying. A number of
observershave also been surprisedat claimsthatthe poverty incidence
in urban areas of Bangladeshovertook rural areas in the mid-1980s.
In additionto the substantiveimportance,these controversiesmake a
good startingpoint for probinginto the methodologyof appliedpoverty
measurement.

Comparabilityof the Surveys

The comparabilityof household surveys over time should always be
of concern in poverty monitoring,though the effects of changes in
survey design are still poorly understood.Past studies for Bangladesh
have largelyignoredpotentiallyimportantdifferencesin HES method-
ology. There was a significantbreak in survey design between the
1981/82and 1983/84surveys. The maindifferencewas the switch from

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
764 EconomicDevelopmentand CulturalChange

a single interview reporting7-day recall of food consumption,to the

use of daily diariesrecordingfood consumptionover the last 24 hours,
spreadover 14 days, with repeatedvisits by interviewers(more inten-
sively to households without a literate member).The numberof food
items identifiedin the questionnairealso increased-roughly doubling
after 1981/82.The new questionnairesfrom 1983/84onwardare likely
to have capturedmore items in the self-consumptioncategory (includ-
ing consumption from common property resources), which is more
importantfor poor households than for others.7 The new survey is
thus likely to yield higher and more accurate estimates of consump-
tion.8 The comparisonsof the 1981/82and 1983/84surveys that show
a (substantial)decrease in poverty incidencecould well be due in large
part to these differences in survey methodology. We shall start our
own investigationwith 1983/84,ignoringthe 1981/82survey. Our dis-
cussions with BBS staff indicatethat the survey methodologyappears
to have stabilizedsince 1983/84.Neither the initialnor finalyears used
for comparisonhere (1983/84and 1991/92)are in any way odd, climati-
cally or otherwise; both were quite close to the trend of agricultural
production(both total and per capita).9

The Food-EnergyIntake Method of Setting Poverty Lines

Even perfectly comparablesurveys do not assure consistent poverty
comparisons.Two of the five studiesreportedin table 1-the estimates
producedby the BBS and the study by Ahmed, Khan, and Sampath-
are based on the food-energyintake (FEI) method of setting poverty
lines; studies for a numberof other countrieshave also followed ver-
sions of this approach.10By this method,the povertyline in each sector
(urbanand rural)and period is obtainedby findingthe expenditure(or
income) level at which the expected value of caloricintake, conditional
on expenditure,equals the predeterminedfood-energyrequirement.
More precisely, let k denote food-energy intake (which is taken
to be a randomvariable),for which the requirementlevel is k' (taken
to be fixed, though this can be readilyrelaxed), and let x denote con-
sumption expenditure.As long as the expected value of food-energy
intakeconditionalon total consumptionexpenditure,E(kIx), is strictly
increasing in x over an interval that includes k', there will exist a
poverty line z such that E(k z) = k'. This method is quite easy to
implement and does not need data on the prices prevailing in each
sector or region.
To illustrate, figure 1 gives estimates of the expected value of
food-energy intake conditional on real food spending for rural and
urban areas of Bangladesh in 1985/86.The regressions reported by
Ahmed, Khan, and Sampathwere adjustedusing our estimates of the
cost of a "basic-needs"food bundleso as to normalizefor differences
in food prices, which tend to be higher in urban areas." We show

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 765

Food-energyintake(KI/person/day)
3,000

2,750
Rural

2,500

2,250

2,000

1,750

1,500
75 80 85 90 95 100 105 110 115 120 125
Realfood expenditure(percentof food povertyline)
FIG.1.-Food-energyintakeandfoodexpenditures, 1985-86
Bangladesh,

that ruralhouseholds have higher caloric intakes at a given real food

expenditure;the ruraldiet is clearly more "calorie intensive"--more
starchy food staples relative to protein, vegetables, sugar, and so on.
The relationshipis also quite flat in both sectors. Thus there is a large
difference in the real food expenditurelevel at which average intake
equals typical requirementlevels. In figure 1 we indicate the implied
real (food) poverty lines for 2,100 calories per person per day; the
expected real expenditureat which an urbanresidentreaches this point
is about 20% higher than it is for a ruralresident. In the version by
BBS, one finds instead the total (food plus nonfood)expenditurelevel
in each sector at which expected caloricintakeequals averagerequire-
ments. The differencein the real value of the impliedpoverty lines is
then likely to increase further, according to differences in nonfood
spendingbehavior.
Why do we observe the shift in the calorie-expenditurerelation-
ship in figure 1?12 A numberof reasons can be suggested. One might
argue that the relativeprice of food is higherin urbanareas (assuming
that food energy has a price elasticity less than unity, as is plausible).
This is questionable, given the high nonfood prices of, for example,
housing in urbanareas." Activity levels for many ruraljobs (notably
agriculturallabor)are higherthanfor manyurbanjobs, althoughpartic-
ipation rates, demographics,and surplus labor may all work in the
opposite direction. It is at least as plausiblethat some nonfood goods
may simply be unavailablein poor ruralareas, precisely because they

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
766 Economic Development and Cultural Change

are so poor, and thus the marketso thin, or because of high transport
costs; the "virtualprice" of those goods may then be so much higher
in ruralareasthat there is a substitutionin favor of food. Better knowl-
edge about nutritionin urban areas may also lead to more balanced
diets, with fewer calories and more micronutrients.
So, for various reasons, people in better-off regions are buying
more expensive calories, reachingfood-energyrequirementsat higher
levels of total spending.'4However, there is nothingin the FEI method
to guaranteethat the ways in which these other factors-such as rela-
tive prices, tastes, and activity levels-influence the poverty lines will
lead one to rank distributionsconsistently with the chosen metric of
household welfare.
More formally,the distributionof caloric intakes can readily vary
between groups such that the regression function E(klx) also varies
systematicallywith the characteristicsof those groups, and there is no
reason to assume that E(k x) rankswelfare levels correctly at a given
value of x. One shouldthen be wary of the poverty lines generatedby
the FEI method, in that people at the poverty line in differentsectors
or years could well have very differentlevels of living by almost any
agreed measure. Indeed, dependingon how these factors vary, it is
quite possible to find that the "richer" sector (by the agreedmetric of
welfare)tends to spend so muchmore on each calorie that it is deemed
to be the "poorer" sector. For example, a case study for Indonesia
found a virtuallyzero rankcorrelationbetween regionalpoverty mea-
sures impliedby FEI-basedpovertylines and measuresthat attempted
to hold constant purchasingpower over basic consumptionneeds.15
Inconsistencies across subgroupsin the real value of the poverty
lines at a given periodalso cloud inferencesover time. Whenthe urban
poverty line has a higher real value than the rural line, and there is
migration from rural to urban areas, the national poverty rate can
increase even though no one is worse off in commandover commodi-
ties. For example, a person in a village living just above the rural
poverty line but below the urbanline who moves to the city and gains
by an amountless than the differencein poverty lines will be counted
as poor after the move.

Defenses of the FEI Method

Against the above arguments,a numberof defenses mightbe made of
the FEI method.
i) There is a sense in which the povertylines based on this method
can be said to be "consistent," namely, that, on average, people at
the poverty line will have the same food-energy intakes relative to
requirements. The issue is whether that constitutes a good basis for
poverty comparisons. It might be if one deemed food-energy intake
(relative to requirements) to be a valid welfare indicator on its own.

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 767

But there appears to be wide agreementthat it is not, even among

exponents of the FEI method of setting poverty lines. In fact, if one
deemed calories to be sufficient, none of this extra work would be
necessary-all one would do is measure caloric shortfallsrelative to
requirements,all of which are alreadyused as data to implementthis
method of setting poverty lines. The FEI method acknowledges (at
least implicitly)that total consumptionof goods and services is a better
welfare indicatorthan is food-energyintake per se.
ii) An argumentsometimes made in favor of this method is that
it reflects differences in preferences between subgroups." Certainly
differencesin preferenceswill affect the poverty lines so derived. But
it is not clear what status one should give those differences when
makingpoverty comparisons. If one group-the urban sector, say-
prefers to consume less food and more clothing at given prices and
real incomes, should one then say they are poorer?Surely not.
iii) Or it might be argued that people in, say, urban areas are
consumingmore expensive calories because the cheaper foods found
in ruralareas are not availablein urbanareas-that they are rationed
in their food choices. In a country such as Bangladesh this seems
unlikely; we cannot think of any food item consumed in rural areas
that could not easily be transportedto urbanareas if the marketwas
there, and (fromcasual observationsonly) it appearsthat pretty much
everything is available in urban areas. (Of course, transport costs
should be factored in; this is uncontentious.)
iv) It is also arguedthat the FEI method is able to reflect other
determinantsof welfare, such as access to publicly provided goods.
But again it is far from clear that the method will do so in a way that
is consistent with defensible normativejudgmentsin makinginterper-
sonal welfare comparisons.For example, access to better health care
and schooling in urban areas than in rural areas may mean that one
tends to consume a diet that is nutritionallybetter balanced-with
relatively fewer calories and more micronutrients.But then the FEI
method will entail a higher poverty line, and so more people will be
deemed poor. This makes little sense.
v) It is possible to use a higherfood-energyrequirementfor rural
areas than for urbanareas and this will go some way toward avoiding
the urbanbias in the FEI method. Requirementswill depend on (inter
alia) the normativejudgments made as to what activity levels are
deemed appropriate.In practice,prevailingmethodsof settingrequire-
ments based on the demographiccomposition and activity levels of
the population may or may not entail higher requirementsin rural
areas.7 And there can be no presumptionthat such refinementsto
the FEI method will achieve a consistent poverty rankingin terms of
commandover consumptionneeds.
vi) Possibly the practical advantagesof the method are compel-

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
768 EconomicDevelopmentand CulturalChange

ling. The methoddoes not requiredata on prices. This is an advantage

in (the surprisinglycommon)cases in which the statisticalagency does
not collect comprehensiveprice data outside the main urban areas.
However, in order to estimatefood-energyintakes one needs data on
the quantities of foods consumed. Typically, expenditures are also
collected, so one can retrieve the "averageprice" or "unit-value"at
the household level. Employing these unit-values can entail further
problems (which we discuss later), but they probably still provide a
better basis for valuationthanthe FEI methodof settingpoverty lines.
So we do not find any of these argumentscompelling. The key
point here is that one wants the method of poverty measurementto
be consistent with the purposeof measurement.When that purpose is
to monitor progress in reducing absolute consumption poverty-
definedin terms of commandover basic consumptionneeds-then the
poverty lines used shouldhave constantvalue in termsof that measure
of welfare. One shouldnot deem a person(livingin a city) who chooses
to buy fewer and more expensive calories as poorerthan anotherper-
son (in a village) when both can afford exactly the same standardof
living. However, the FEI method does not pass this test. That is not
a problem if one is setting a single poverty line (for the country as a
whole at one date, say), or if one is not interestedin makingpoverty
comparisonsacross groups. But in other applicationsit can be a seri-
ous problem.

The Cost-of-Basic-NeedsMethod
The main alternativemethodof setting poverty lines found in practice
is the cost-of-basic-needs (CBN) method; a version of this method
was used by the remainingstudies in table 1. The method dates from
Seebohm Rowntree's study of poverty in York around 1900.'8By this
approachthe poverty line is set as the cost in each sector and at each
date of a normative "basic needs" bundle of goods. That bundle is
typically chosen to be sufficient to reach the predeterminedcaloric
requirement,with a compositionthat is consistent with the consump-
tion behaviorof the poor.1
This method has its problemstoo. The poverty lines it generates
can be interpretedas Laspeyres cost-of-living numbers.20So utility-
compensatedsubstitutioneffects in consumptionare ignored.21While
the (implicit)bundleof goods in the FEI methodalmostcertainlyvaries
too much to be consistent with the same standardof living, the (ex-
plicit) bundlein the CBN methodvariestoo little. Difficultiesin setting
nonfood "basic needs" and in valuing their cost at local prices are
also a problemfor this method, as we discuss below.
However, these problems are less worrisome in our view than
those with the FEI method, and it is also easier to test the sensitivity
of the results to alternativeways of dealingwith the problemsin the

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 769

CBN method. The most compelling argumentin favor of the CBN

method for making poverty comparisonsis that it explicitly aims to
control for differences in purchasingpower over basic consumption
needs, while the FEI method does not. The CBN method can at least
claim to provide a first-orderapproximationof what we are trying to
measure.
We use the CBN method here to update the estimates in table 1
to 1991/92,and to revise the series on a consistent basis. Our aim is
not to derive a definitive series of poverty estimates for Bangladesh.
Rather, it is to obtain a reasonably defensible alternative series to
those in table 1, which can be implementedwith the availabledata, and
to test the robustnessof the key qualitativeresultsto our measurement
assumptions.The following discussion describes the main differences
found among past studies for Bangladesh that have used the CBN
method, and our own choices.
A relativelyuncontroversialissue in the Bangladeshliteratureus-
ing the CBN method has been the food items included in the "mini-
mum consumptionbundle";the bundlewe have used is given in table
2; with minor variations this is the same as originally used by M.
Alamgir, among others.22This correspondsto an average per capita
daily intake of 2,112 calories and 58 grams of protein, in accordance
with the Food and AgricultureOrganizationrecommendedstandard
for South Asian countries, based on the notion of balanced nutrition
accordingto the age, sex, and occupationalcompositionof the popula-
tion.23However, views may reasonablydifferon the appropriatecalo-
ric requirement,reflecting,for example, differentviews of what consti-
tutes an acceptable activity level-itself a normativejudgment.24
A more significantsource of discrepancy among earlier poverty
estimates is probablythe choice of prices used to convert the norma-
tive minimumfood consumptionbundle into monetarypoverty lines.
Urban retail prices for food items in the consumption bundle have
been used in past work-with the exception of the work of Mahabub
Hossain and Binayak Sen-with an assumed fixed discount rate to
obtain a ruralretail price level.25A movement in urbanprices out of
line with ruralprices may then generatepotentiallylarge errors in the
estimatedpoverty measures. The use of the CPI for updatingthe base
year poverty line may generate errors in the poverty trends since the
construction of the CPI (based on 113 goods) includes many items
that clearly fall outside the typical consumptionbundle of the poor in
Bangladesh. An alternative source of price informationis the set of
implicitunit-valuesfor food in the HES. The implicitprices are derived
by dividing reported expendituresby quantities for each food item.
These give the actual expenditures on a unit of consumptionpaid in
each sector and date, and so they reflect the underlyingdifferences in
prices. However, that is not all they reflect. Differences in the quality

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
 TABLE 2
RURAL AND URBAN CONSUMERUNIT VALUES

PER CAPITA
NORMATIVEDAILY MEAN RURAL CONSUMERUNIT
REQUIREMENT VALUES CALCULATEDFROMHES (Tk
ITEMSINCLUDEDIN THE
MINIMUMCONSUMPTION BUNDLE Calories Grams 1983/84 1985/86 1988/89 19

Rice 1,386 397 7.52 8.00 10.01

Wheat 139 40 5.42 6.11 7.63
Pulses (masur and khesari) 153 40 7.48 13.11 16.12
Milk (cow) 39 58 6.66 8.92 9.95
Oil (mustard) 180 20 37.08 41.71 40.30 5
0 Meat (beef) 14 12 23.88 34.97 36.44 4
Fish (fresh water) 51 48 18.31 22.74 23.21 2
Potato 26 27 3.27 4.17 6.69
Other vegetables (leafy and non-
leafy) 36 150 3.04 3.34 4.33
Sugar (gur) 82 20 9.71 11.57 14.73
Fruits (banana) 6 20 5.42 7.40 9.21
Total (calorie/g) 2,112 832 ........
Poverty line expenditure on food
(Tk/person/day) ... ... 6.64 7.80 9.15
Rural CPI for nonfood ... ... 100.00 122.44 150.40 2
Urban CPI for nonfood (industrial
workers) ... ... .. ..

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 771

of the consumptionswithineach categoryin the survey data can entail

shifts in unit-valueseven if prices for given qualitiesare not different.
Thus there are pros and cons to each of these methods. We shall use
the implicitfood prices in the HES to determinethe cost of the norma-
tive minimumdiet in each sector and year to obtain the food compo-
nent of the povertyline. We will, however, considerthe possible biases
inherentin this method. Like most household surveys, the HES does
not provide quantities consumed of nonfood goods (only expendi-
tures);so the implicitprices cannotbe derived.We thus have no option
but to updatethe cost of the nonfood componentfor 1983/84over time
using the nonfood componentof the CPI as calculatedusing base-year
weights.26
Setting the nonfood component of the poverty line is a further
potential source of contention, since there is no agreed anchor analo-
gous to the role played by food-energyrequirementsin settingthe food
componentof the poverty line. And some commonpractices-such as
using the meanfood shareof the poor in each subgroupor period-are
questionable,in that they can imply far more generousallowances for
nonfood goods in richer subgroupsor periods.27 Against this one may
want to allow for differences in the price of food relative to nonfood
goods across sectors or regions. Unfortunately, without believable
spatial price data on nonfood goods (allowing for the heterogeneity
in quality of housing, e.g.) it is difficult to make such adjustments
convincingly.28For example, the ruralpoverty estimates by Hossain
and Sen set the allowance for nonfood goods at 30%of that for food,
while Sen and Islam's urban series used a markupof 40%; it is not
clear how much of this difference can be attributedto differences in
real levels of livingbetween the two sectors and tastes (differencesthat
one would not want to introduceinto the method)versus differencesin
relative prices (which one would want to incorporate).Atiq Rahman
and TrinaHaque used a constant markupacross the two sectors, set-
ting the nonfood allowance at 25% of the food poverty line. This is
defensible if one believes that relative price differences between the
two sectors are negligible (or preferences are homothetic, which is
unlikely),29although25%appearsto be a rathermeagerallowance for
nonfood goods even for rural areas.30 An appropriatelyspecified de-
mandmodel could be used to help resolve this issue,31 but this requires
access to the unit-record(household-level)data, which so far has not
been possible for Bangladesh. For this study we decided to set the
nonfood allowance at 35%of the food poverty line in 1983/84,splitting
the differencebetween the Hossain-Sen ruralallowance and the Sen-
Islam urbanallowance.32We will test robustnessto this choice.
From the foregoingdiscussion it is evident that the main ingredi-
ents of a poverty measure-the caloricrequirement,the food bundleto
achieve that requirement,andthe allowancefor nonfoodgoods-entail

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
772 Economic Development and Cultural Change

normativejudgments.Thatcannotbe avoided, althoughsome methods

appear to be more consistent and defensible than others in the judg-
ments made. However, once the relativitiesof poverty lines over time
and space are agreed on (though, as we have noted, this is itself a
contentious step), these normativejudgmentscan be collapsed to the
choice of the absolute level of the poverty line at a given place and
date. The sensitivity of the resulting poverty comparisons to those
judgmentscan then be tested by applyingrecent results in the applica-
tion of stochasticdominancetests to distributionalcomparisons.33This
allows us to isolate to what extent (if any) these normativejudgments
alter key qualitativeconclusions. That is the approachwe follow here.

Poverty Measures
There has also been a degree of controversyover how the information
on poverty lines and the distributionof consumption expenditures
should be aggregatedin the form of a poverty measure. Unlike some
of the issues discussed above, this issue has received a great deal of
attention.34The three poverty measuresused in this study attempt to
capturethree aspects of poverty:its incidence, its depth, and its sever-
ity. The specific measuresfollow.
i) The head-countindex (H), given by the percentageof the popu-
lation living in households with a consumptionper capita that is less
than the poverty line. This can be interpretedas a measure of the
"incidence" of poverty. The measurehas the advantagethat it is easy
to interpret, but it tells us nothing about the depth or severity of
poverty.35
ii) The poverty-gapindex (PG), definedby the mean distance be-
low the poverty line as a proportionof that line (where the mean is
formed over the entire population, counting the nonpoor as having
zero poverty gap).36One can interpretthis as a measure of poverty
"depth." Its disadvantageis thatit is unaffectedby changesin inequal-
ity among the poor.
iii) The squared poverty-gap index (SPG) of James Foster, J.
Greer,and ErikThorbecke,definedas the meanof the squaredpropor-
tionate poverty gaps (againthe mean is formedover the entire popula-
tion, counting the nonpoor as having zero poverty gap).37Thus the
poverty gaps are weighted in aggregation,with greater weight given
to larger gaps, and where the weights are simply the poverty gaps
themselves. This simplechangeto the conventionalpoverty-gapindex
allows the index to reflect changes in the "severity" of poverty, in
that it will be sensitive to inequalityamong the poor.
All three measures (like almost all measures found in practice)
are functions of both the mean consumption (4) of each subgroup
normalized by the poverty line (z), and the Lorenz curve for the distri-
bution of consumption. So we can write the poverty measures in the

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 773

generic form P(?i/z, L), where L denotes all the parametersof the
Lorenz curve. The variable i/z is simply the purchasingpower of
mean consumptionin termsof the poverty bundleof goods; for brevity
we will call it "mean real consumption,"but it shouldnot be forgotten
that it is "real" in terms of this specific bundle of goods. The effects
of changes in pt/zon poverty, holdingthe Lorenz curve constant, and
of certain changes in the Lorenz curve can be predictedby looking at
the statisticalpropertiesof the initial distribution.38

III. AlternativePovertyMeasuresfor Bangladesh,1983-92

Our Estimates
Table 3 gives our alternativeestimates based on the version of the
CBN method describedabove. Comparingthe end points, 1983/84and
1991/92, every poverty measure in both urban and rural sectors fell
(table 3). But progresswas uneven duringthe period. Indeed, all three
of our measures show an increase in poverty nationallyafter 1985/86,
in contrast to the BBS results based on the FEI method. Most of the
increase was due to rising poverty in the ruralsector; that sector saw
steadily worsening poverty measuresfrom 1985/86on. This holds for
all three measures. The sizable increase in the squared poverty-gap
index in rural areas that started in the mid-1980s suggests that the
poorest of the poor, as well as those nearthe poverty line (as indicated
by the less dramaticrise in the head-countindex), were also suffering
fallingliving standards.The overall dropin poverty incidencebetween

TABLE 3
POVERTY MEASURES FOR BANGLADESH

Squared
Head-Count Poverty-Gap Poverty-Gap
Index (%) Index (%) Index (x 100)
Urban:
1983/84 40.9 11.4 4.4
1985/86 30.8 7.3 2.5
1988/89 35.9 8.7 2.8
1991/92 33.6 8.4 2.8
Rural:
1983/84 53.8 15.0 5.9
1985/86 45.9 10.9 3.6
1988/89 49.7 13.1 4.8
1991/92 52.9 14.6 5.6
National:
1983/84 52.3 14.5 5.7
1985/86 43.9 10.4 3.5
1988/89 47.8 12.5 4.6
1991/92 49.7 13.6 5.1
NOTE.-The ruralpopulationshares are 88.7%(1983/84),
87.2%(1985/86),86.6%(1988/89),and 83.4%(1991/92).

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
774 Economic Development and Cultural Change

the end points was not sufficientto prevent rising numbersof poor;
the nationalhead-countratiofell at 0.6%per year compoundedover the
entire period, implyingthat the total numberof poor increasedat about
1.5%-2%per year.
All three poverty measuresindicategreaterpoverty in ruralareas
than urban areas. It appears that, for Bangladesh, the BBS's and
Ahmed, Khan, and Sampath'sfindingsof higherpoverty rates in urban
compared with rural areas (table 1) are due to the aforementioned
urbanand ruraldifferencesin the real value of the poverty lines gener-
ated by the FEI method.39In terms of commandover basic consump-
tion needs, ruralareas experiencegreaterincidence, depth, and sever-
ity of poverty. It also appearsthat the shiftingreal value of the BBS
poverty lines (underlyingthe results reproducedin table 1) has caused
a reversal in the direction of change in measured poverty between
1985/86and 1988/89;the BBS methodshows a decrease in both sectors
while our results show an increase.
There are some markedquantitativedifferences between our re-
sults and those in table 1. Suppose, however, that we discount esti-
mates based on the FEI method(namely,2 and 3 in table 1)andexclude
the 1981/82survey on comparabilitygrounds. Then we are left with
agreementon the directionof change in poverty across the studies in
table 1 and our estimatesin table 3 in those cases for which a compari-
son is possible. In particular,there is agreementthat the incidence of
poverty fell in both urbanand ruralareas between 1983/84and 1985/
86, and that it rose in both sectors between 1985/86and 1988/89.At
least for such qualitativecomparisons,the differences in method ap-
pear to matterlittle. In the rest of this section we will attemptto assess
how robust these results are to changes in measurementassumptions.

Robustness to Biases in Using Unit-Values as Prices

As noted in Section II, unit-values(expenditureper unit quantityfor
each of the survey's food categories)reflect qualitychoices as well as
prices. It is reasonableto assume that qualityis a normalgood, so the
existence of quality biases implies that the real value of the poverty
line will be an increasingfunctionof the real mean. However, as long
as the income effect on demand for quality is not too high-
specifically, that the expenditure-shareweighted average elasticity of
unit-valueto total expenditureis less thanunity-the poverty measure
will still be a strictly decreasingfunction of the real mean. Assuming
that mean real expenditureis higherin urbanareas, it is clear that this
source of bias would lead us to overestimateurbanpoverty measures
relative to ruralmeasures, but we will get the rankingright. We will
also tend to underestimate the impact on poverty of changes in the
real value of mean consumption, as at least some of the change will

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 775

be dissipated in shifts in the real value of the poverty line associated

with changes in food quality. We would thus suspect a dampeningof
fluctuationsin absolutepoverty measuresover time, but not a reversal
in orderings.

Robustness of the Urban-RuralPoverty Comparisonto Differences

in Relative Prices
Because the same normativeconsumptionbundleis used in both urban
and rural areas, the CBN method that we have used does not adjust
the poverty lines for differencesin relativeprices. Would such adjust-
ments alter our key qualitative conclusions? As table 2 shows, the
relative unit-values of the food items are differentin urban and rural
areas. The ruralfood poverty line in 1991/92is 87%of the urbanline,
but the ratio differs from 72%(for milk) to 99%(for oil). Among food
goods, our use of a fixed food bundle means that we have overesti-
matedthe true urban-ruralcost-of-livingdifference,since urbanhouse-
holds will be able to exploit the relativeprice differencesthroughsub-
stitution without being worse off. This means that we have
overestimatedurbanpoverty relativeto ruralpoverty. Our qualitative
conclusion that poverty is higherin ruralareas does not change.
The picture is less clear when we consider nonfood goods. We
have used the same proportionatemarkupfor nonfood goods in both
sectors. This is not obviously wrong. While it is plainthat many goods
are more expensive in urban areas (notably food and housing), it is
not known whether the cost of food needs relative to nonfood needs
of the poor is different.If the relativeprice of nonfood goods is higher
in urbanareas then we will have underestimatedurbanpoverty relative
to ruralpoverty. Since we do not have nonfood prices, this is untest-
able. But what we can do is ask, How much of an errorin our estimate
of the relative cost of nonfood goods to the urban poor would be
needed to reverse our qualitativeconclusion that poverty is higher in
rural areas? For 1991/92we calculated the urban poverty line that
would be needed to equalizethe urbanhead-countindex with the rural
index for that year of 52.9%. (This was done by line search, stopping
when the percentagedeemedpoor was equalto the firstdecimalplace.)
We found that the urbanallowance for nonfood goods would have to
be 85%of the food poverty line before it ceased to be true that poverty
incidence in ruralareas is higherthan it is in urbanareas. As long as
the relative price of nonfood goods in urbanareas is less than double
that in ruralareas the sector rankingfor 1991/92will be preserved. For
the other two poverty measures, the critical allowance for nonfood
goods needed to equalize poverty measures across sectors was 75%
for PG and 73%for SPG. The poverty rankingfor these measures is
thus slightly less robust than for the head-count index. However, it

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
776 EconomicDevelopmentand CulturalChange

remainstrue that a considerablyhigherrelativeprice of nonfood goods

in urbanareas would be needed to reverse our poverty orderingof the
two sectors.
Of course, since this issue arisespreciselybecause we do not have
spatial data on nonfood prices, it is difficultto go much further. We
cannotrule out the possibilitythat we have underestimatedthe relative
price of nonfood basic needs for the poor in urbanareas, though the
underestimationwould have to be very substantialto reverse our quali-
tative conclusion.

Robustness to Choice of Poverty Line and Measure

As we have emphasized,there is, inevitably, a degree of arbitrariness
in setting any poverty line and in choosing a poverty measure, and it
is importantto ask how robust the above results are to those choices.
Figure 2 gives the head-count indexes in 1983/84and 1991/92for a
wide rangeof potentialpoverty lines, expressed as a percentageof the
benchmarkpoverty line we have used in table 3. By varyingthe pov-
erty line one traces a poverty incidence curve (PIC), which is simply
the empiricalcumulativedistributionfunction. Results from the appli-
cation of stochastic dominancetheory to poverty comparisonscan be
used here.40 One can readily show that if the PIC for distributionA
lies above that for B up to some maximumpoverty line, then the claim
that poverty is higher in A is robust to the choice of poverty line up
to that maximum.It is also robust to the choice of poverty measure

Cumulativepercentof people
100
90
80
70 +
60
50
40
30 Urban 1983/84
20 + +Urban 1991/92
X Rural 1983/84
10-
0 0-r- Rural;1991/92
'-
20 40 60 80 100 120 140 160 180 200
Percent of the poverty line
FIG. 2.-Poverty incidencecurves for urbanand ruralBangladesh,
1983/84and 1991/92.

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 777

within a broadclass, includingboth the poverty-gapindex and squared

poverty-gapindex, as well as a wide rangeof other measures.41 Figure
2 shows that the reductionin urbanpoverty over the period is robust,
as is the conclusion that poverty is higherin ruralareas than in urban
areas. On the other hand, the comparisonof ruralpoverty in 1983/84
with 1991/92is quite ambiguous;some poverty lines and measuresgive
different results than others, though at no point is there more than a
small (undertwo percentagepoints) shift either way in the rural PIC
over the period.
To test dominance within subperiodswe divide the period into
two: 1983/84to 1985/86,and 1985/86to 1991/92,for which figures 3
and 4 give the correspondingPICs. For the first subperiodwe findthat
the conclusion that poverty fell is robustto the choice of poverty line
and poverty measure, as is the conclusion that poverty is higher in
ruralareas duringeach period. The conclusion that poverty increased
in both sectors from 1985/86to 1991/92is also robust to the choice of
line and measure, and againthis is true of the conclusion that poverty
is greaterin ruralareas.

Consistency with Wage Data

Other poverty data may provide a useful check. The rural poor in
Bangladeshdepend heavily on supplyingagriculturallabor and so the
real agriculturalwage rate is an importantdeterminantof their wel-
fare.42It has been claimed that real wages have been on a generally

Cumulativepercent of people
100
90-
80-
70-
60-
50-
40
30-
20-

10
- -
20 Urban 1983/84 ? Urban 1985/86
0 Rural 1983?84 -Rural 1985/86
20 40 60 80 100 120 140 160 180 200 220
Percent of the poverty line
FIG.3.-Poverty incidencecurvesfor 1983/84and 1985/86

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
778 EconomicDevelopmentand CulturalChange

Cumulative percent of people

100
90 -
80
70-
60-
50
40-
30-
20-
Urban 1985/86 Urban 1991/92
10- 20-
0 X Rural 1985/86 Rural 1991/92
20 40 60 80 100 120 140 160 180 200 220
Percent of the poverty line
FIG.4.-Poverty incidencecurvesfor 1985/86and 1991/92

rising trend since the beginningof the 1980s,continuingto at least the

early 1990s.43If true, this would suggest either that our result of rising
ruralpoverty since the mid-1980sis questionableor that other determi-
nants of poverty have militatedagainstthe rising real wage effect.
However, there is conflictingevidence on what has been happen-
ing to wages for agriculturallabor in Bangladesh. There are two
sources within BBS for data on agriculturalwages, the Agricultural
Statistical Wing (ASW) and the National Income Wing (NIW).44The
ASW data had not been collected for a numberof years since 1989.45
The extent of discrepancybetween the two sources for years in which
both were collected is striking,particularlysince 1985;figure 5 gives
estimates of real wages from both sources, as well as an independent
(but less regular)series from village surveys by the BangladeshInsti-
tute of Development Studies (BIDS).46The increase in the real wage
indicated by the NIW series over the latter half of the 1980s is not
confirmedby the ASW data. The discrepancybetween the two sources
has been increasingover time. In the early 1980s they were roughly
equal; by July 1993, the NIW wage estimate was 50%higherthan the
ASW estimate.47 The BIDS estimates are in closer accord with the
ASW numbers.
The question arises why the NIW data are so out of step with
these other sources. The ASW agriculturalwage survey was intended
to be representativeof ruralBangladeshas a whole. It covers all 23
districts, using weekly surveys of employers at thana-levelruralcen-

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 779

Real wage rate in Taka per day, 1981/82 prices

26-
24 )NIW
22
- NIW ASW
20-

18-

16 - BIDS
ASW
14

12
1981 1983 1985 1987 1989 1991 1993 1995

wagesinBangladesh.
FIG.5.-Real agricultural Sources:NationalIncome
Wing(NIW)andAgricultural StatisticsWing(ASW)of Bangladesh
Bureauof
Statistics,andBangladesh
Instituteof Development
Studies(BIDS).

ters, and the survey appears to be based on a representativesample

for estimatingthe average agriculturalwage in Bangladesh.48By con-
trast, the NIW series is far more limited in both coverage and scope.
The NIW datawere initiallycollected for six regionalcenters, namely,
Dhaka, Chittagong, Rajshahi, Rangpur, Khulna, and Sylhet. Since
1987/88the national average is based on four centers only (with the
exclusion of Rangpurand Sylhet). The NIW data are just for one
thana-level center in each of four divisions and the selection of the
thana is restricted to a rural area adjacent to the (urban)divisional
center. For example, Demra (a well-developedthana adjacentto the
capital) is chosen as the reference point for the collection of NIW
weekly wage data for the Dhakadivision. Also, the numberof respon-
dents interviewedis quite small in the case of the NIW data; only two
respondents per thana are interviewed. Comparedto that, the ASW
survey currentlyadministersa muchmore elaboratelystructuredques-
tionnaireto three randomlyselected wage-employersper union in each
of the 464 thanas of the country.49 The difference between the two
data sources could well reflectin parta higherrate of economic growth
in urban and near-urbanrural areas compared with growth in more
typical ruralareas.
The above observations lead us to conclude that the ASW data
can be considered a more reliableguide to average wages for agricul-
tural labor in Bangladesh.The ASW data indicate an increase in the
real wage up to the mid-1980s,but a decline thereafter.Thus both sets

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
780 EconomicDevelopmentand CulturalChange

of data and our reanalysis of the household survey data suggest that
the signs of progressin the first half of the 1980swere not sustained.

IV. Some Implications

Proximate Causes of the Changes over Time
Assessments of the causative factors underlyingmeasuredchanges in
poverty can also dependgreatlyon measurementmethods. For exam-
ple, if the poverty line is changingin real value as some function of
the mean of the distributionthen this will bias estimates of how much
impact growth and contractionin average living standardswill have
on absolute poverty; indeed, when the elasticity of the poverty line to
the mean reaches one, the poverty measurewill depend only on how
relative inequalitieschange; everyone could be better-offbut nothing
will happento measuredpoverty. By aimingto arriveat poverty mea-
sures that are consistent across time and space in how they treat a
given standardof living, we also hope to be able to throw a more
powerful light on the proximatecauses of changes in living standards
of the poor.
Table 4 gives furtherstatistics that can help in understandingthe
results in table 3 and in figures 2 and 3. The following observations
can be made.
i) The fourth column of table 4 gives our estimate of the basic-
needs purchasingpower of meanconsumption,that is, mean consump-
tion as a percentageof the poverty line. There was a gain in the pur-
chasing power of mean consumptionbetween 1983/84 and 1985/86,
and a decline after that date. The changes in L/z track quite closely
the changes in poverty measuresin table 3; typically, when the mean
rose (fell), poverty fell (rose). For the other estimates in table 1, the
impactof changes in the mean on poverty is more obscure;even when
the differingestimates agree on direction,the response varies greatly

TABLE 4
SUMMARY STATISTICS ON GROWTH AND INEQUALITY

Mean Gini
Poverty Line Consumption Mean/Poverty Index
(Tk/Person/Month) (Tk/Person/Month) Line (%) (%)
Urban:
1983/84 301.72 396.53 131 29.8
1985/86 368.62 581.13 158 31.4
1988/89 453.65 695.19 153 32.6
1991/92 534.99 817.12 153 31.9
Rural:
1983/84 268.92 284.84 106 24.6
1985/86 319.06 373.93 117 24.6
1988/89 379.08 435.39 115 26.5
1991/92 469.13 509.67 109 25.5

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 781

(e.g., the 10%growthin the ruralp/z from 1983/84to 1985/86produced

between a 5.5% and 17.4%drop in the ruralhead-countindex across
the estimates in table 1). This diversityreflects the differencesin mea-
surement methods, and, as we have argued, some of those methods
are hard to defend.
ii) There was a markedsectoral imbalancein the rate of growth
in the real value of mean consumption.While p/z increased by about
16%in urbanareas between 1983/84and 1991/92, it rose by only 3%
in ruralareas.
iii) The last column of table 4 gives the Gini indexes by year and
sector. There was an increase in relative inequalitieswithin both sec-
tors up to 1988/89,with some improvementafter that.
A casual inspection of the summarystatistics in table 4 thus sug-
gests that the growth and contractionin real consumptionper capita
may well have been an importantfactor in the changes in poverty
measures observed in table 3. But a second "proximate"cause of the
slow progressin reducingpovertynationallycan be suggested,namely,
the rise in inequalityin both sectors over much of the period. A com-
parison with table 1 suggests, however, that other methods can imply
quite different conclusions. For example, despite a decline between
1985/86and 1988/89in the basic-needspurchasingpower of the rural
mean, and a rise in inequality, the BBS estimates based on the FEI
method indicate a fall in poverty incidence.
Quantificationof the relative importanceof growth and distribu-
tional change to poverty measures over this period is possible using
some simple counterfactualexperiments.
Experiment1. How much did the intrasectoralinequalityin the
benefits of growthin the 1980scontributeto the slow progress in pov-
erty reduction?To addressthis question, we can simulatethe poverty
measures if the Lorenz curve had not changed over the period; thus
we comparethe actual poverty measurefor the base data, P[(Q/Z)83/4,
L83/4], withthe simulatedmeasure, L83/4].50 We give the
P[(•/z)91/2,
results in table 5. If all consumptiongroups had sharedequally in the
growth that occurredin urbanareas (leavingthe Lorenz curve in that
sector unchanged)then the head-countindex of poverty would have
droppedby 10.8 percentagepoints over the period, instead of 7.3. In
the ruralsector, the head-countindex would have fallen by 2.4 points,
instead of 0.9. Both the poverty-gapindexes would also have fallen
furtherif not for the increase in inequality.
Experiment2. How much furtherwould poverty have increased
if not for the growth that did occur? Here we need to simulate the
poverty measures for 1991/92 that would have been observed if the
mean had been that of 1983/84(keeping the actual Lorenz curve of
1991/92), that is, to compare P[(pC/z)83/4,L83/4] with the simulated mea-
sure,
P[(p•/z)83/4,L91/2]. This simulation is also given in table 5. All

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
782 EconomicDevelopmentand CulturalChange

TABLE 5
SIMULATED
POVERTY
MEASURES

Squared
Head-Count Poverty-Gap Poverty-Gap
Index (%) Index (%) Index (x 100)
Urban:
1983/84(actual) 40.9 11.4 4.4
1991/92(actual) 33.6 8.4 2.8
1991/92(with no changein inequality) 30.1 7.5 2.6
1991/92(with no growthin p/z) 44.5 12.7 4.9
Rural:
1983/84 (actual) 53.8 15.0 5.9
1991/92(actual) 52.9 14.6 5.6
1991/92 (with no change in inequality) 51.3 14.0 5.4
1991/92 (with no growth in p/z) 55.4 15.6 6.1
1991/92 (with urban rate of growth in
R/z, 1984/84-1991/92) 41.0 10.3 3.6

poverty measureswould have been higherby the end of the period if

not for the (even modest)growth;in urbanareas, an extra 3.6%of the
populationwould have been poor without growth; in rural areas, an
extra 1.6%.
Experiment 3. How much did the regional imbalance in the
growth of mean real consumptioncontributeto the slow overall prog-
ress? Suppose that mean consumptionrelative to the poverty line in
ruralareas had grown at the same rate as the urban sector; by 1991/
92 mean consumptionin ruralareas would have been Tk 577.5, about
13%higher than its actual value. Here we simulatethe poverty mea-
sures for ruralareas in 1991/92using the actual Lorenz curve for that
sector and year but replacingthe actualmean with that impliedby the
urbanrate of growth. The ruralhead-countindex would have fallen a
further10percentagepoints, entailinga 24%reductionover the period,
instead of the 2% actually observed; PG would have fallen by 31%
(instead of 3%);and SPG would have fallen by 39% (instead of 5%;
table 5).

Rates of Poverty Reduction underAlternativeGrowthPaths

How responsive are Bangladesh'spoverty measures likely to be to
futuregrowth in the basic needs purchasingpower of mean consump-
tion? The answer will depend of course on the extent to which the
poor share in that growth, that is, on what contemporaneouschanges
in inequalityoccur duringthe growthprocess. Table 6 gives two sets
of estimatesof the elasticitiesof the threepoverty measuresto changes
in mean consumption,descriptionsof which follow.
1. Inequitablegrowth elasticities are those implied by the actual
changes in the poverty measuresand the mean relative to the poverty

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 783

TABLE 6
GROWTH ELASTICITIES UNDER ALTERNATIVE ASSUMPTIONS

ELASTICITY
OFPOVERTY
WITHRESPECT
TOMEANCONSUMPTION

Squared
Head-Count Poverty-Gap Poverty-Gap
Index Index Index
Urban:
Inequitablegrowth - 1.1 -1.6 -2.2
Neutralgrowth -2.1 -3.0 -3.9
Equitablegrowth:
Sector neutral -2.3 -3.6 -4.9
Ruralbiased -2.2 -3.3 -4.4
Rural:
Inequitable growth -.6 -.9 - 1.7
Neutralgrowth -1.8 -2.6 -3.2
Equitablegrowth:
Sector neutral -1.8 -2.9 -3.7
Ruralbiased -1.9 -3.1 -4.2

line, as observed over the period 1983/84to 1991/92.51 So these elastici-

ties reflect the effects of rising inequality over the period in each
sector.
2. Neutral growth elasticities are those that assume that relative
inequalitiesdo not change; table 7 gives these elasticities.52Since the
poor experiencethe same rate of growthas othergroups,these elastici-
ties will be higher (in absolute value) than for case 1.
3. Equitable growth elasticities are anchored to the pattern of
growthobservedin Indonesiaduringthe 1980s,wherebyrobustgrowth

TABLE 7
POINT ELASTICITIES OF THE POVERTY MEASURES IN TABLE 3

TOTHEMEAN
ELASTICITY ELASTICITY
TOTHEGINIINDEX

Head- Squared Head- Squared

Count Poverty- Poverty- Count Poverty- Poverty-
Index Gap Index Gap Index Index Gap Index Gap Index
Urban:
1983/84 -1.9 -2.6 -3.2 .6 2.1 3.6
1985/86 -2.2 -3.2 -4.0 1.3 3.4 5.4
1988/89 - 1.9 -3.2 -4.3 1.0 3.2 5.4
1991/92 -2.1 -3.0 -3.9 1.1 3.1 5.1
Rural:
1983/84 -1.8 -2.6 -3.1 .1 1.2 2.3
1985/86 -2.3 -3.2 -4.0 .4 1.7 3.0
1988/89 - 1.9 -2.8 -3.4 .3 1.6 2.8
1991/92 -1.8 -2.6 -3.2 .2 1.3 2.5

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
784 EconomicDevelopmentand CulturalChange

in mean consumptionwas combinedwith a modest rate of decline in

inequality.Two cases are considered:sector neutralassumes that the
elasticity of the Gini index to the mean is - 0.2 in both sectors, while
ruralbiased assumes an elasticity of -0.1 in urbanareas and -0.4 in
ruralareas; these assumptionsheld duringthe period 1984-87.53
It can be seen that the switchfromthe growthprocess of the latter
half of the 1980sto a distribution-neutral growthprocess would entail
a markedincreasein the elasticitiesof the poverty measuresto growth.
The (absolute)elasticity of the ruralhead-countindex to growthin the
mean would rise from 0.6 to 1.8. A 2% rate of growth in mean con-
sumption in that sector would bring the ruralpoverty rate down by
1.2%per year under the inequitablegrowth path. The rate of decline
in the poverty rate would be 3.6%per year if that growthprocess was
sufficientlybroad to allow consumptionsof the poor to grow at the
same rate as those of others. The differences are not as marked in
urbanareas, and they are somewhatless pronouncedfor the poverty-
gap indexes.
The head-count indexes will respond to distribution-neutral
changes in the mean with an elasticity of about 2 in both sectors.
The elasticities of the poverty-gapindexes are approximately3-4. The
elasticities to the mean tend to be slightlyhigher(in absolute value) in
urbanareas than in ruralareas (for any given combinationof year and
measure),and they are higherfor SPG than for PG, and lowest for H.
Thus we findthat Bangladesh'spoverty measuresare likely to be quite
responsive to change in the mean, holdingthe Lorenz curve constant,
and that this is even more true of the distribution-sensitivemeasure
(which better reflects the circumstancesof the poorest) than of the
simple head-countindex.
The switch to either "equitablegrowth" path would have little
impact on the growth elasticity of the head-count index, given that
Bangladesh's poverty rate is so high (in particular,that the mean is
close to the poverty line). But that switch would give a boost to the
growthelasticitiesof the poverty-gapindexes, particularlySPG, which
rises to an elasticity of 4.9 in urbanareas and 3.7 in ruralareas even
in the sector-neutralcase (4.5 in the rural-biasedcase).
The higher the growth elasticity, the lower the actual rate of
growthneeded to reducethe numberof poor. Table8 gives the implied
estimates of the minimumrate of growth in nationalincome needed
before the total numberof poor in Bangladeshwould begin to fall. The
additional assumption made in deriving these estimates is that the
share of nationalincome that is consumedby householdsremainscon-
stant.54We find that the switch to either a neutralor equitablegrowth
path will entail a sizable drop in the minimum rate of growth needed to
prevent rising numbers of poor; at Bangladesh's projected population
growth rate of 2% per year in the 1990s, the minimum growth rate in

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 785

TABLE8
MINIMUMRATES OF GROWTHIN
NATIONALINCOMETO PREVENT
RISING NUMBERSOF POOR

POPULATION
GROWTHRATE

2.5% 2.0%
Inequitablegrowth 6.1 4.9
Neutralgrowth 3.9 3.1
Equitablegrowth:
Sector neutral 3.8 3.1
Ruralbiased 3.8 3.1
NOTE.-The nationalelasticities in
each case are givenby the poverty-share
weighted means of the urbanand rural
elasticitiesin table 6; the weightedelas-
ticities are -0.7, - 1.8, - 1.9, and - 1.9
for each of the four growth processes,
respectively.

GNP per capita needed to prevent rising numbers of poor switches

from 5%per year underthe inequitablegrowthpathto about 3%under
the neutralor equitablegrowthpaths. With a nationalincome growth
rate of approximately4%per year (typicalof the 1980s)such a switch
would make the differencebetween risingandfallingnumbersof poor.

V. Conclusions
Poverty measurement requires value judgments and behavioral as-
sumptionsto help interpretthe available-and invariablyimperfect-
data. That will always be true. The more interestingquestion is how
much bearingthe choices made by the analysthave on the key conclu-
sions drawn. This case study has illustratedthat those choices will
sometimes, but not always, matter to qualitative conclusions about
progress in reducing poverty, and the sectoral structureof poverty.
Credible poverty monitoringcalls for considerable care in critically
evaluatingdata and methods.
A survey of recent estimates of poverty measuresfor Bangladesh
suggests some worrying discrepancies, even when the estimates are
based on the same survey data set and use a very similarspecification
of food-energyrequirements.Past studies have come to differentcon-
clusions aboutthe directionsof changein poverty over time, and about
which of the sectors, the urbanor the rural, is poorer. The only way
of resolvingthe issue is to look closely at the data and at the normative
and behavioralfoundationsof the methods used to analyze it, in the
hope that defensiblecriteriacan be foundfor rejectingsome estimates
in favor of others.

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
786 EconomicDevelopmentand CulturalChange

We find that all of the recent estimates are questionable from

one point of view or another, though some of the problemsare more
worrisomethan others. We have arguedagainstthe common practice
of definingthe poverty line as the consumptionexpenditureor income
level at which the expected food-energyintake equals preset require-
ments in each sector, region, or time period. Althoughnot ideal, we
have arguedin favor of an alternativemethod found in the literature
in which the poverty line is an estimateof the cost of an explicit bundle
of basic consumptionneeds. Applyinga version of the latter method,
we have proposedour own revised series, tryingto preservethe better
featuresof past work, while rejectingothers. Withbetterdataon prices
and better access to the underlyinghousehold-level data one could
improveon our estimates; we are hopefulthat will happenin the near
future. However, fromthe data that are available,there is a good case
for believing that our estimates are more consistent over time and
space than some others found in the literature--"consistent" in the
sense that two persons with the same commandover basic consump-
tion needs will be treated the same way. Such consistency is a value
judgment, but a defensible one when a purpose of poverty measure-
ment is to inform our understandingof how the economy and policy
influencelevels of living of the poor.
We find that the fall in poverty in the early 1980s reported by
some observers appears to largely reflect changes in measurement
methodology, including survey design. Differences over time and
space in the real value of the poverty lines used in some studies also
raise serious doubtsabouttheirconclusions. On removingwhat appear
to be the mainmethodologicalproblemsin past estimates, we findthat
there was a reductionin poverty incidence, depth, and severity around
the mid-1980s, but that it was not sustained after that. We also find
that the claims by some past studies that in Bangladeshurbanpoverty
incidence has overtakenruralpoverty incidence are questionable.All
our poverty measures are higher in rural areas, and this has not
changedover the period. Measuredconsistently, it appearsthat prog-
ress has also been uneven across sectors; indeed, almost the entire
modest drop in poverty over the whole period is attributableto gains
to the urbanpoor.
What is at stake in aiming for sound measurementmethods is
more than the reliabilityof poverty assessmentsover time, thoughthat
is clearly important.Unreliablemethods can also confound attempts
to assess the causative factors at work, includingpolicy changes. The
methods we have used here can also help throwlight on the proximate
causes of changes in poverty measures, and possible implicationsfor
development policy.
It might be of some comfort if we could conclude that the slow
and uneven progress in reducing poverty in Bangladesh was "simply"

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 787

a problem of too little growth;but what we have seen in Bangladesh

is generally low growth combinedwith signs of risinginequality, sug-
gesting that the patternof growthhas not been particularlyequitable.
If not for the rising inequalityover the period, we would have seen
lower poverty measuresin 1991/92;the urbanpoverty rate would have
been 10%lower than it actually was in 1991/92, while the rural rate
would have been 3% lower. The sectoral imbalance in progress in
reducing poverty is also worrisome. Inequality between urban and
rural sectors has increased; while the urban mean real consumption
was 24% higher than the rural rate in 1983/84, the urban mean was
40%higherin 1991/92.If the basic-needspurchasingpower of average
expendituresin the ruralsector had been able to grow at the same rate
as that in the urbansector, then we would have seen the ruralpoverty
rate reduced by one quarterover this period, instead of the modest
2% drop actually observed.
Provided that future economic growth is not associated with a
worseningof existing inequalities,Bangladesh'spoverty measures can
be expected to respondquite elastically.The outcomewill then depend
largelyon just how muchreal growthis achieved. However, a continu-
ation of the 1980s trend of rising inequality in both urban and rural
areas will make it considerablymore difficultto preventrisingabsolute
numbers of poor in Bangladesh. For example, at population growth
rates around2%-2.5%per year, nationalincome would have to grow
at about 3%-4%per year to preventan increase in the numberof poor
if inequality neither increases nor decreases; that rate is consistent
with recent performance.However, the recent pattern of rising in-
equality in both urbanand ruralareas would mean that a growth rate
in national income of 5%-6% per year would be needed before the
numberof poor stops rising.

Notes
* Theviewsexpressedinthisarticleareoursandshouldnotbe attributed
to the WorldBank.Wehavebenefitedfromthe commentsof HafezGhanem,
EmmanuelJimenez,HaiderAli Khan,MichaelLipton,PradeepMitra,Peter
Nicholas,ShekharShah,Dominique vande Walle,thejournal'sreferees,and
seminarparticipantsat the WorldBankandparticipants at a meetingof the
Associationfor EconomicandDevelopment Studieson Bangladesh.Shahadat
HossainandTaherUddinAhmedat the Bangladesh Bureauof Statisticshave
also helpedus on mattersrelatedto the datausedhere.
1. BangladeshBureauof Statistics (BBS), Report on the Household Ex-
penditure Survey, 1988-89 (Dhaka: BangladeshBureau of Statistics, 1991).
in Proceedingsof
Also see TaherUddinAhmed,"Povertyin Bangladesh,"
the Workshopof Disseminationof CurrentStatistics (Dhaka:BangladeshBu-
reauof Statistics,1991).
2. Fora surveyof theevidencesee MichaelLiptonandMartinRavallion,
"Poverty and Policy," in Handbookof DevelopmentEconomics, vol. 3, ed.
JereBehrmanandT. N. Srinivasan
(Amsterdam:
North-Holland,
1995).

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
788 Economic Development and Cultural Change

3. Abu Abdullah, "Introduction,"Bangladesh DevelopmentStudies 18

(1990):iii-vii; AzizurRahmanKhan, "Povertyin Bangladesh:A Consequence
of and a Constraintof Growth,"BangladeshDevelopmentStudies 18 (1990):
19-34; SiddiqurOsmani,"StructuralChangeand Poverty in Bangladesh:The
Case of a False TurningPoint," BangladeshDevelopmentStudies 18 (1990):
55-74, and "Notes on Some Recent Estimates of Rural Poverty in Bangla-
desh," BangladeshDevelopmentStudies 18 (1990):75-87; MartinRavallion,
"The ChallengingArithmeticof Poverty in Bangladesh,"BangladeshDevel-
opment Studies 18 (1990):35-54.
4. MartinRavallion, "AgriculturalWages in Bangladesh:What Do the
Figures Really Show?" Journalof DevelopmentStudies 31 (December 1994):
334-45.
5. Mustafa K. Mujeri, Quazi Shahabuddin,and Salehuddin Ahmed,
"MacroeconomicPerformance,StructuralAdjustmentsand Equity:A Frame-
workfor Analysisof Macro-MicroTransmissionMechanismsin Bangladesh,"
in MonitoringAdjustmentand Poverty in Bangladesh, Reporton the Frame-
work Project, CIRDAPStudy Series no. 160 (Dhaka,Bangladesh:Centrefor
IntegratedRuralDevelopmentfor Asia and the Pacific, 1993).
6. Ravallion,"The ChallengingArithmeticof Poverty."
7. Khan, "Poverty in Bangladesh."
8. There are other differencesbetween the 1981/82survey and the last
four, includingthe change to a smallersample size, althoughthis is unlikely
to matterin the present context.
9. Of the fourHES surveyyears, only 1988/89was unusualin this respect,
because of severe flooding.The agriculturaloutput data referredto are from
various issues of the MonthlyStatisticalBulletinof Bangladesh.
10. V. M. Dandekarand N. Rath,Poverty in India (Pune:IndianSchool
of Political Economy, 1971);Azizur RahmanKhan, "Poverty and Inequality
in Rural Bangladesh,"in Poverty and Landlessness in Rural Asia (Geneva:
InternationalLabourOrganisation[ILO], 1977);SiddiqurOsmani,Economic
Inequality and Group Welfare (Oxford:Oxford University Press, 1982); J.
Greer and Erik Thorbecke, "A Methodologyfor MeasuringFood Poverty
Applied to Kenya," Journal of DevelopmentEconomics 24 (1986): 59-74;
Satya Paul, "A Modelof Constructingthe PovertyLine," Journalof Develop-
ment Economics30 (1989):129-44; Aly Ercelawn,"AbsolutePoverty as Risk
of Hunger:Norms, Incidence, and Intensity for Ruraland Urban Pakistan"
(Applied Economics Research Centre, University of Karachi, 1991, mimeo-
graphed);SudhirAnandand ChristopherHarris, "Issues in the Measurement
of Undernutrition,"in Nutritionand Poverty, ed. S. R. Osmani(Oxford:Ox-
ford UniversityPress, 1992).
11. AkhterU. Ahmed,HaiderA. Khan,andRajanK. Sampath,"Poverty
in Bangladesh: Measurement,Decomposition and IntertemporalCompari-
son," Journalof DevelopmentStudies 27 (1991):48-63. Later we discuss the
food bundlewe have used and how it is valued. Note that Ahmed, Khan, and
Sampathgive regressions of food energy intake against (log) nominal food
expenditures. Using our estimates of the food poverty line (the local cost
of an appropriatefood bundle), one can use their regressions to derive the
relationshipbetween food energy intakeand real food expenditures(this sim-
ply requiresan adjustmentto their intercept).
12. This is not peculiarto Bangladesh,andwe suspectit is quitecommon.
See, e.g., Martin Ravallion and Benu Bidani, "How Robust Is a Poverty
Profile?" WorldBank EconomicReview 8 (1994):75-102, on Indonesia.
13. To some extent this reflectsdifferencesin housingquality,but that is
not the only factor;groundrents are clearlyhigherin urbanareas. We do not

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 789

know of evidence for Bangladesh,but on hedonic-pricedifferentialsfor hous-

ing in Java see MartinRavallionand Dominiquevan de Walle, "Urban-Rural
Cost-of-LivingDifferentialsin a Developing Economy," Journal of Urban
Economics 29 (1991): 113-27.
14. This is not a theoreticalpropertyof the method,and empiricalexcep-
tions exist. For example, GauravDatt, "Povertyin India, 1950-1991"(Policy
ResearchDepartment,WorldBank, Washington,D.C., 1995,mimeographed),
finds that the poverty lines for urbanand ruralIndia generatedby the food-
energy method appearto have constant real value in terms of commodities
consumedby the poor.
15. Ravallionand Bidani.
16. Greerand Thorbecke.
17. Some of the activitiesin the urbaninformalsector may actuallyentail
higher energy requirementscomparedwith activities in rural areas. Casual
observationsuggests that activities such as rickshawpullingand brick break-
ing, which representa majorsource of the urbanpoor's employment,would
entail similarlyhigh energy expenditurescomparedwith agriculturallabor.
18. Anthony B. Atkinson, TheEconomicsof Inequality(Oxford:Oxford
University Press, 1975).
19. Some applicationsof this method have set the composition of that
bundle to achieve nutritionalrequirementsat minimumcost given prevailing
prices. This could easily entail a diet that is alien to the tastes of poor people.
We do not follow or recommendthatapproach.Rather,we recognizeexplicitly
that the reference bundle of goods is a normative standard,which can be
chosen to be consistent with the actualconsumptionchoices of any reference
group of households.
20. Dividing throughoutby some reference poverty line one obtains a
Laspeyres price index.
21. Though certain non-utility-basedconcepts of "welfare" would not
deem this to be a problem;for furtherdiscussion see Ravallionand Bidani.
22. M. Alamgir, "Some Analysis of Distributionof Income, Consump-
tion, Savingand Poverty in Bangladesh,"BangladeshDevelopmentStudies 2
(1974): 737-818; Qazi KholiquzzamanAhmad and MahabubHossain, "An
Evaluationof Selected Policies and Programmesfor Alleviationof RuralPov-
erty in Bangladesh"(BangladeshInstitute of Development Studies, Dhaka,
1984, mimeographed);M. Muqtada, "Poverty and Inequality:Trends and
Causes," in Bangladesh-Selected Issues in Employmentand Development,
ed. Rizwanul Islam and M. Muqtada(New Delhi: ILO-ARTEP, 1986), pp.
75-92; Atiq RahmanandTrinaHaque, "Povertyand Inequalityin Bangladesh
in the Eighties:An Analysisof Some Recent Evidence," ResearchReportno.
91 (BangladeshInstitute of Development Studies, Dhaka, 1988); Mahabub
Hossain and BinayakSen, "RuralPovertyin Bangladesh:TrendsandDetermi-
nants," Asian DevelopmentReview 10 (1992): 1-34; Binayak Sen and Quazi
TowfiqulIslam, "MonitoringAdjustmentand Urban Poverty in Bangladesh:
Issues, Dimensions, Tendencies," in MonitoringAdjustmentand Poverty in
Bangladesh (n. 5 above). The precise compositionof the bundlehas differed,
sometimesto accommodatedifferencesin price data, thoughoften the choices
madewere not particularlyappropriatefor the poor;for examples, see Hossain
and Sen. The precise requirementhas also varied slightly, withinthe range of
2,100-2,200 calories.
23. Some observers consider this standardto be on the high side for
Bangladesh.See Khan, "Poverty and Inequality"(n. 10 above).
24. Anandand Harris(n. 10 above); Lipton and Ravallion(n. 2 above).
25. Hossain and Sen; Rahmanand Haque.

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
790 Economic Development and Cultural Change

26. Alternativelyone can fix the food share over time (such as in the
Orshanskymethod, used for updatingthe officialpoverty lines for the United
States; see Isabel Sawhill, "Poverty in the U.S.: Why Is It So Persistent?"
Journal of Economic Literature26 [1988]:1073-1119).This does not require
data on nonfoodprices. The problem,however, is that the price of food rela-
tive to nonfood goods may change over time, creatingdrift in the real value
of the poverty line, and hence inconsistencies in the poverty comparisons
drawn. This may be importantin Bangladeshover this period, since the rela-
tive price of food staples was generallyfalling.
27. MartinRavallion,Poverty Comparisons,Fundamentalsof Pure and
Applied Economics, vol. 56 (Chur,Switzerland:HarwoodAcademic, 1994).
28. The only attemptwe know of is the articleby Ravallionand van de
Walle (n. 13above), which used hedonicregressionsto estimatea spatialprice
index for housingappropriateto the poor in Java. This requiresgood data on
dwelling characteristics.
29. Of course food prices are typicallyhigherin urbanareas than in rural
areas (shown in table 2), but many nonfoodprices (such as housing, although
this accounts for only a small share of expenditureby the poor) are higherin
urbanareas. And it is the differencein relativeprices that mattershere.
30. Hossain and Sen.
31. Ravallion,Poverty Comparisons.
32. Withthe fallingrelativepriceof food, the allowancefor nonfoodneeds
increases over the period, risingto 42%in 1991/92.
33. AnthonyB. Atkinson,"Onthe Measurementof Poverty,"Economet-
rica 55 (1987):749-64; James Foster and A. F. Shorrocks, "Poverty Order-
ings," Econometrica56 (1988): 173-77; Ravallion,Poverty Comparisons.
34. For a survey of this literature, see James Foster, "On Economic
Poverty: A Survey of Aggregate Measures," Advances in Econometrics 3
(1984):215-51.
35. The limitationsof the head-countindex as a measureof poverty are
now widely appreciated,particularlyfollowing AmartyaSen, "Poverty: An
OrdinalApproachto Measurement,"Econometrica46 (1976): 437-46, and
Poverty and Famines: An Essay on Entitlementand Deprivation (Oxford:
OxfordUniversityPress, 1981),and an extensive elaborationis not called for
here. For a recent survey of the issues and referencessee Ravallion,Poverty
Comparisons.
36. See JamesFoster, J. Greer,and ErikThorbecke,"A Class of Decom-
posable Poverty Measures,"Econometrica52 (1984):761-66, for a definition
of the poverty-gapindex, which has advantagesover the income-gapratio,
obtainedwhen the mean is only formedover those who are poor; for further
discussion see Ravallion,Poverty Comparisons.
37. Foster, Greer, and Thorbecke.
38. Nanak Kakwani, "Povertyand Economic Growthwith Applications
to C6te D'Ivoire," Review of Income and Wealth39 (1993):121-39. Also see
S. M. Ravi Kanbur, "Measurementand Alleviationof Poverty," IMF Staff
Papers 34 (1987):60-85.
39. Ahmed, Khan, and Sampath(n. 11 above).
40. Atkinson, "On the Measurementof Poverty."
41. For further discussion see Ravallion, Poverty Comparisons(n. 27
above).
42. On the poverty profileby landholdingclass in Bangladeshsee Martin
Ravallionand Binayak Sen, "Impactson RuralPoverty of Land-Contingent
Targeting:Some Further Results for Bangladesh," WorldDevelopment 22
(1994): 823-38. The importanceof the real wage rate to the rural poor in

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
MartinRavallionand Binayak Sen 791

Bangladeshhas been a prominentlink in analyses of the causes of the 1974

famine. See Sen, Poverty and Famines; MartinRavallion,Marketsand Fam-
ines (Oxford:OxfordUniversityPress, 1987).Althoughthereis no comparable
time-series evidence for Bangladesh,the strong link between rural poverty
measuresand the real agriculturalwage rate in Indiais borneout by the results
of MartinRavallionand GauravDatt, "Growth, Wages and Poverty: Time
Series Evidence for India," Policy Research WorkingPaper (World Bank,
Washington,D.C., 1994).
43. See RichardPalmer-Jones,"An ErrorCorrected?And Whatthe Fig-
ures Really Show," Journalof DevelopmentStudies 31 (1994): 346-51.
44. It is not always obvious which source is being used in appliedwork;
Palmer-Jones,e.g., simplysplices the two withoutdistinction.Publicationsof
the BBS have not been particularlytransparenteither.In the successive statis-
tical yearbooks(1980, 1986),usually only "BBS" is mentionedas the source
of the ASW data, which appears under the heading "region-wise wage of
agriculturallabor." Data from the NIW, which are also publishedin the year-
book under the title "daily wage rates of selected groups of workers," also
give only "BBS" as the source. So it is difficult,at least at first glance, to tell
which source is used. The 1992yearbookhas addedto the confusionby attrib-
uting both to "price section, BBS," a section within NIW.
45. Note that the ASW series is not availablefor several time points:
January-November1990;January-December1992;January-June1993.Data
were not collectedfor these periods.Availabilityof ASW datain the published
form is even less: the 1993statisticalyearbookprovidesdistrict-wisedata up
to 1987/88;the recently published1992yearbookof agriculturalstatistics pre-
sents data for 1986-89, 1990(December),and 1991.
46. For details on the sources and deflatorsused in constructingfig. 5,
see MartinRavallionand BinayakSen, "New Evidenceon AgriculturalWages
in Bangladesh"(Policy ResearchDepartment,WorldBank,WashingtonD.C.,
1994, mimeographed).
47. A composite series was constructedby Palmer-Jones.The composite
series splices the NIW series onto the ASW numbersat 1985;doing so gives
the impressionof a substantialgain in real wages, as shown in fig. 5 if one
tracksthe ASW series up to 1985and then switches to the NIW numbersafter
that.
48. The survey was at thanalevel before July 1993,but at the finerunion
level from then on.
49. Before July 1993, ASW covered all 23 greater(old) districts and all
464 thanas, using weekly surveys of four or five preselected employers at
thana-levelruralcenters.
50. This was estimated using parameterizedLorenz curves. We tested
both the beta and generalizedquadraticspecifications;both gave valid Lorenz
curves on these data, thoughthe beta specificationgave a better fit. The ex-
plicit formulasfor all three poverty measuresas functions of the parameters
of both Lorenz curves can be found in GauravDatt and MartinRavallion,
"Growthand RedistributionComponentsof Changesin Poverty Measures:A
Decompositionwith Applicationsto Brazil and India in the 1980s," Journal
of Development Economics 38 (1992): 275-95. A user-friendlyprogramfor
PCs, POVCAL, that implementsthese simulationsis availablefrom Martin
Ravallion. See Shaohua Chen, Gaurav Datt, and MartinRavallion, "POV-
CAL: A User-Friendly Computer Program for Poverty Analysis Using
Grouped Data" (Policy Research Department, World Bank, Washington,
D.C., 1992).
51. So, e.g., the elasticity of the urbanhead-countindex of - 1.1 is the

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions
792 Economic Development and Cultural Change

ratio of the percentagedecrease in that index over the period (- 17.85, from
table 2) to the percentageincrease in the mean, normalizedby the poverty
line (16.79, from table 3).
52. The formulasfor these elasticities can be found in Kakwani (n. 38
above). While we focus solely on the 1991/92distribution,the results in table
7 indicate that these elasticitiesare quite stable over time.
53. MartinRavallionand Monika Huppi, "MeasuringChanges in Pov-
erty: A MethodologicalCase Study of Indonesia duringan AdjustmentPe-
riod," WorldBank Economic Review 5 (1991): 57-84. Over this period in
Indonesia, the nationalGini index fell by 3.0% (from 33.1%to 32.1%)while
mean real consumptionper personrose by 15.7%.In urbanareas the Gini fell
by 1.2%,while the meanrose by 12.1%,while in ruralareasthe corresponding
percentageswere 5.5 and 14.6.
54. The numberof poor will then remainconstant if the rate of growth
in nationalincome equals the rate of populationgrowthtimes 1 - 1/(growth
elasticity).

The Journal of Modern History

The Journalof ModernHistorycombinestraditional and innovative
methodsof historicalscholarship,
examining crucial topicsandassessing
recenttrendsin historiography.Thejournal'sgeographical andtemporal
sweep-the historyof the Europeancontinent,including Russiaandthe
Balkans,since the Renaissance--makesituniqueinitsfield. As the
premierreviewing journalforEuropeanhistory,itpresentsseveralmajor
reviewarticlesand upto sixtyshorterreviewseach issue. Ifyour
researchrequiresthe fusionof social,political,intellectual,
andeconomic
history,the JMHwillenhanceyourscholarlywork.
Editors
JohnW.BoyerandJuliusKirshner,
One year rates: Individuals$35.00; Institutions$82.00; AHAIndiv.Members$27.00;
Students (withcopy of ID)$27.00. Outside USA add $7.00 for postage. Canadians,
include7% GST. Visa and MasterCardaccepted. To subscribe, send check or
complete chargecard information to The Universityof ChicagoPress, JournalsDivision,
Dept.SS6SA, P.O. Box 37005, Chicago,IL60637.
2/96

This content downloaded on Fri, 18 Jan 2013 00:14:56 AM

All use subject to JSTOR Terms and Conditions