<<

. 4
( 13)



>>

(see Klasen, 2000, 2007a, for a discussion). This therefore allows an assessment
of intra-household poverty dynamics which is impossible using income data
but might be quite important for studying intra-household inequalities (see
also Haddad and Kanbur, 1990). This advantage, which is already present in a
static assessment of poverty using a capability/functioning framework, might
be easily extended within a dynamic framework, as we illustrate below.


4.2.2. Limitations
Apart from stating some advantages of an extension of chronic poverty to
non-income dimensions, it is also important to name some problems. Prob-
ably the most important objection to such an extension is that it would not
yield very useful new information as many non-income dimensions of well-
being do not change much over time. Moreover, change in some non-income
measures usually means improvements, at least in the way it is measured. The
most extreme example of this would be to use years of schooling to track edu-
cation poverty of individual adults. This indicator is likely to stay the same for
the vast majority of adults once they have left the educational system and if it
changes, it will only go up, but never down (as surveys usually only keep track
of improvements in education, but not of loss of knowledge/skills over time).
Thus many non-income measures of well-being seem to exhibit a great
deal of inertia and most non-income poor would be chronically non-income
poor and there would be no point in distinguishing between them and the
small number of transitory non-income poor (see McKay and Lawson, 2003).
In contrast, the evidence on income poverty is that there is a great deal of
churning, and in many countries most of the poor at any point in time are
transitory poor while a smaller share of households are chronically poor (see
Baulch and Hoddinott, 2000).
There are several possible replies to this objection. First, to the extent that
these non-income measures adequately re¬‚ect the functioning shortfall in
question, the inertia in these measures correctly suggests that in many devel-
oping countries many people are chronically deprived of critical functionings.
For example, those adults (of whom many are female) in developing countries



81
Isabel Günther and Stephan Klasen

that never had the chance to be schooled will be chronically educationally
poor. Stating this might be obvious but from a well-being perspective we
need to occasionally remind ourselves that all attempts to achieve universal
enrolments for children will do nothing to combat education poverty among
adults. In these cases it is particularly interesting to see whether these chron-
ically non-income poor are also chronically income poor and how the two
measures are related in a static and dynamic context.
Second, some indicators of measuring non-income well-being achievements
are not adequately re¬‚ecting the functioning in question. For example, adult
height only re¬‚ects the nutritional status during the phase of growing up,
but not the current one. Also, years of schooling say nothing about the
level of functional education a person has at a point in time. To track this
we would need different measures such as test scores and functional literacy
and numeracy surveys which only exist in some countries (e.g. OECD, 2000).
These scores are likely to move more over time and can go up or down for
adults.
Third, despite the fact that a well-being indicator for an individual is not
changing over time, it may sometimes be useful to consider a household
perspective. For example, while being educated oneself is clearly valuable in
and of itself, sometimes there are also individual bene¬ts of education of
other household members. To the extent this is the case, it might be useful
to consider the average education of household members or possibly even
the highest education level existing in the household (see Basu and Forster,
1998). These indicators will clearly move more over time than individuals™
educational level.
Fourth, for some indicators there is considerable movement for children,
but little (or no) movement among adults (e.g. years of schooling). Thus
it might be useful to separately track changes in non-income poverty of
children and adults which will generate different insights. Lastly, we show
in section 4.4 that there are a range of indicators where there is quite a lot
of dynamics over time so that there indeed is an empirical justi¬cation to
examine chronic (and transitory) poverty in a non-income perspective.
A second objection is that current survey instruments lack the tools to
systematically track poverty in the non-income dimension. There is clearly
a valid point as many surveys do not systematically track, for example, the
health or nutrition status of all individuals across time using comparable
measures. In the survey that we use, there are also shortcomings in this
respect. However, this objection can only lead to efforts to improve survey
design rather than abandon this interesting approach.
A third objection is that it is dif¬cult to interpret the linkages between
income and non-income poverty dynamics for two reasons. One is the differ-
ing magnitude of measurement error in income and non-income dimensions
of poverty which might make it dif¬cult to interpret differences in chronic



82
Measuring Chronic Non-Income Poverty

income vs. non-income poverty. While this is an important issue which we
also discuss below, it focuses attention on the role of measurement error in
the assessment of chronic poverty and a comparison between chronic income
and non-income poverty might actually help shed light on this important
issue. 6 The second problem of interpretation deals with the fact that the
income indicator will always refer to households, whose size and composition
might have changed over time, thereby affecting the poverty status of that
household. In contrast, the non-income assessment of poverty will usually
focus on individuals that are present in both periods. While one indeed has
to bear in mind this difference when household chronic poverty is compared
with individual non-income chronic poverty, one can just as easily use the
existing household boundaries to calculate non-income household poverty
(as we do below).
A last objection is that when measuring non-income poverty dynamics
several new conceptual questions arise. For example, what is education and
health poverty among children? How can one de¬ne such poverty to be
chronic or persistent? Am I education poor only if I am out of school?
Or also if I am falling behind in progressing through school? Or if I also
have a worsening performance? When do I become chronically education
poor? Similarly, is stunting already an indicator of chronic poverty since it
is related to persistent lower than required energy intake (UNICEF, 1998) or
is only persistent stunting an issue? Clearly these are serious questions and
below we explore the empirical impact of some choices of answers relating
to these dif¬cult questions. But also here, the call is for more work extending
the concept of chronic poverty to these cases rather than to abandon the
enterprise.
Thus we believe that it is well worth studying chronic non-income poverty
and the approach we take here in this chapter is to simply explore whether,
given data and measurement constraints, reasonable ways to conceptualize
and measure non-income poverty can be extended across time and whether
they will generate useful additional information about static and dynamic
aspects of well-being.



4.3. Methodology

4.3.1. Measurement of chronic poverty
To measure chronic poverty two methods have been proposed: the ˜spells™
(McKay and Lawson, 2003) and ˜component™ (Jalan and Ravallion, 1996)
approach. The ˜spells™ approach de¬nes households as chronically poor who
6
Also, in longer panels, it would be possible to control for this problem through appropri-
ate econometric techniques.




83
Isabel Günther and Stephan Klasen

have always been poor, i.e. whose per capita household consumption has been
below the poverty line in all observed points in time. The transient poor are
those who have only temporarily been poor. In contrast, the ˜component™
approach distinguishes permanent (average) consumption of a household
from temporary variations in household consumption. Hence, whereas the
˜spells™ approach classi¬es households as either chronic poor or transient poor
the ˜component™ approach calculates the ˜chronic™ and ˜transient™ component
of households™ poverty and a classi¬cation of households into chronic and
transient poor households is not possible.
In this study we opt for the ˜spells™ approach as we only have a two-wave
panel at hand. We de¬ne individuals to be chronically poor in the non-
income dimension if they are poor in both periods considered. Those who
are poor in either period but not chronically poor are thus the transitory
poor and those being poor in neither period are de¬ned as the non-poor. In
a two-wave panel it is dif¬cult to assess whether observed transient poverty is
caused by ¬‚uctuating welfare indicators or whether transient poverty is caused
by individuals falling into poverty or escaping from poverty, i.e. we cannot
say whether we observe stochastic or structural changes in the well-being of
individuals.
Similar to the income dimension, we will de¬ne ˜poverty lines™ for the non-
income dimensions based on reasonable (but essentially arbitrary) notions of
who should be considered as poor in the relevant dimension (see below for
details). Also, we will, in line with the literature on chronic income poverty,
treat poverty in the income and non-income dimensions as a dichotomous
yes/no question and thus will not consider depth or severity of poverty. 7 In
addition, as both from a theoretical perspective as well as from a measurement
perspective non-income poverty for adults and children is often de¬ned and
measured differently and thus should show different dynamics, we further-
more analyse poverty dynamics for these two subgroups of the population
separately.



4.3.2. Indicators of non-income poverty
The question arises which non-income indicators should be analysed.
Whereas for a theoretical discussion of temporary and long-term well-being
an analysis of a very broad range of functionings might be appropriate,
when undertaking empirical studies it should be more useful to focus on a
smaller subset of basic functionings. We therefore focus on education and

7
Clearly, considering depth and severity of income and non-income poverty and con-
sidering the correlations between income and non-income dimensions in them would yield
additional useful information and should be taken into account in further work.




84
Measuring Chronic Non-Income Poverty

health (approximated with the nutritional status of individuals 8 ), since these
are probably some of the most critical and commonly agreed capabilities
(Hulme and McKay, 2005). These non-income indicators have the additional
advantage that they are measured at the individual level in contrast to e.g.
housing or service access which can (as income) only assess chronic poverty
of households.
For children who are below the age of 18 years we use stunting as an indi-
cator of health or nutritional deprivation, whereas for adults of 18 years and
older we use the body mass index (BMI). Moderate (severe) under-nutrition
(or nutrition poverty) is de¬ned as being below a z-score of ’2 for children or
being below a BMI of 18.5 for adults.
The z-score is calculated as the height for a child minus the median height
of a reference standard (of children of the same age), divided by the standard
deviation of that reference standard. The reference standard used is the com-
monly used US-based reference standard recommended for use by the World
Health Organization (WHO) for monitoring under-nutrition everywhere since
1987 (see Klasen, 2007b, for further details).
While most analysts agree that the z-score is particularly accurate in measur-
ing nutrition problems of children below the age of 6, there are questions with
regard to its applicability to populations outside the USA for children older
than the age of 6, as growth after 6 seems to differ even in well-nourished
populations across the world (see WHO, 1995, for a discussion). Thus one
should view the application of this indicator until the age of 18 with some
caution.
One should also note that this measure of anthropometric shortfall essen-
tially makes a probabilistic assessment of the likelihood that a child is under-
nourished. As a result, some well-nourished children might be wrongly clas-
si¬ed as under-nourished because they have genetically short parents while
other children might wrongly be classi¬ed as well nourished even though
they are under-nourished but this does not show up in their height due to
their genetically tall parents. Thus we expect some noise in these anthropo-
metric data. However, while this noise will affect static assessments of under-
nutrition, it should not seriously affect the dynamics of under-nutrition.
For adults, height is not an indicator of current nutritional status, and the
BMI as a measurement for under-nutrition is thus chosen instead. The BMI is
de¬ned as the weight in kilograms divided by the height squared in metres
of individuals. While a low BMI is surely an indicator of severe nutritional
problems, the precise cut-off is controversial. Also, due to secular changes in

8
In an earlier version, we also considered a morbidity indicator, but this measure only
captured very recent illnesses and not a more general health status and is therefore not very
suited for an analysis of ˜health™ poverty. Clearly, this is an issue that could be solved by
including more detailed health questions in household surveys. See Schultz (2002, 2003) for
a discussion of particularly useful health indicators.




85
Isabel Günther and Stephan Klasen

dietary patterns and exercise in developing countries, malnourished people
might have an adequate BMI or even show up as overweight, but still lack
important nutrients and access to healthy foods. Thus some ˜health™ poor
might not be captured using this indicator. 9
Moderate (severe) education poverty for adults of 16+ years of age is de¬ned
as having less than nine (four) years of education. Moderate education poverty
for children of 6“15 years of age is de¬ned as being out of school within the
¬rst nine (four) years of education. Four years of education refer to completed
primary school. Nine years of education refer to completed lower secondary
school. In addition, we do not only consider children who have been in
school in one observation period but not in school in the other observation
period as transient poor, but also those children who were in school in both
observation years n and n + t, but did not complete t years of schooling during
the observation period, are considered as transient poor.
Clearly, while the choice of the schooling variable seems defensible for
adults, the choice is somewhat arbitrary for children. One could equally well
consider only those children who are out of school in the two observation
periods as poor as well as consider all children who are behind in their
educational programme, taking account of their age (in either years), as edu-
cational poor. This would include children who are behind in their education
programme already in the ¬rst observation period as well as children who fall
behind during the observation period, i.e. children who progress less than the
number of years between the two waves of the panel. All of these problems
could be circumvented by the use of educational test scores, but hardly any
household surveys, let alone panel surveys, collect such data on a regular basis.
Examining several non-income indicators, which is inevitable when study-
ing well-being from a functioning/capability perspective, the question of an
appropriate aggregation and weighting arises if one wants to generate sum-
mary measures of well-being (e.g. Atkinson and Bourguignon, 2000; Ramos
and Silber, 2005). Alternatively, one can simply report the individual function-
ings/capabilities without weighting and aggregating them, thus generating
partial orderings of well-being outcomes.
In this study, we opt for the latter approach and did not calculate a
composite indicator but examine chronic and transitory deprivation in these
indicators separately. In addition to the usual problems that emerge when
aggregating and weighting different non-income indicators, it is particularly
dif¬cult to interpret such a composite measure in a dynamic perspective, as
different non-income indicators show quite different dynamics. For exam-
ple, education poverty using our indicators is largely irreversible as people
have reached adulthood, while nutrition poverty can be reversed as general
conditions improve. Moreover, when analysing multidimensional poverty in

9
See, for example, Henderson (2005) for a discussion.




86
Measuring Chronic Non-Income Poverty

a dynamic perspective not only the aggregation and weighting of different
non-income indicators becomes an issue but also how to do this over
time. 10


4.3.3. Research questions
Applying the described non-income indicators to the study of chronic under-
development, we ¬rst analyse if and to what extent income and non-income
indicators show the same poverty dynamics. We study both the level of
chronic (and transient) non-income and income poverty as well as the cor-
relation of income and non-income poverty dynamics. The ¬rst analysis
assesses from a macro-perspective whether the same share of individuals suffer
from chronic income and non-income poverty whereas the latter approach
analyses from a micro-perspective if the same individuals would be identi¬ed
as chronically poor whether income or non-income indicators are used.
In a second step we study individual poverty dynamics within households,
which includes an analysis of the differences between individual and house-
hold poverty dynamics, as well as intergenerational poverty dynamics, which
analyses the persistence of poverty of different generations living in the same
household. Such an analysis of individual chronic poverty is usually not
possible with income indicators.
Last, we might also de¬ne chronic poverty as multidimensional poverty,
i.e. examine the number of dimensions of deprivation (including income and
non-income dimensions) at a point in time and over time.



4.4. Data

The data we use are the Vietnam Living Standard Survey (VLSS), which is a
two-wave panel conducted in 1992/3 and 1997/8. The ¬rst round comprises a
sample of 4,799 (23,838) and the second round a sample of 5,999 (28,509)
households (individuals). Four thousand three hundred and ¬ve of these
households were interviewed in both years, which allows to track 17,829 indi-
viduals over a ¬ve-year time period. As we limited our analysis to households
and individuals that were present in both years, there might be a problem
of attrition bias in the sense that the households and individuals studied
do not fully represent the population. However, simple probits indicate that
the attrition bias in the VLSS is quite low, i.e. basically random (Baulch and
Masset, 2003).
10
For example, should an individual who is poor in one dimension in the ¬rst period but
not so in the latter, but who is not poor in a second dimension in the ¬rst period, but poor in
the second period, be considered as chronic poor (deprived of one non-income dimension in
either period) or as transient poor (altering deprivation)?




87
Isabel Günther and Stephan Klasen

Also children below the age of 5 years in 1997 are excluded from the sample
as they were not yet born in 1992. For comparison with non-income poverty
dynamics we also calculate income poverty dynamics. We de¬ne moderate
poverty as per capita household consumption below the of¬cial poverty line
and severe poverty as per capita household consumption below the of¬cial
food poverty line. The of¬cial (food) poverty line, which is provided by the
General Statistical Of¬ce in Vietnam, is 1.160.000 (750.000) Vietnamese Dong
for 1992 and 1.790.000 (1.287.000) Vietnamese Dong for 1997, respectively.
Note that in this study we use per capita household consumption. 11



4.5. Empirical Results

4.5.1. Level of chronic poverty
Table 4.1 shows the extent of chronic and transient poverty measured with
income and non-income indicators. Depending on the measures we use (and
whether we observe adults or children) we come to quite different conclusions
about poverty dynamics and the level of chronic poverty in Vietnam.
In general one can however state that nutritional and particularly educa-
tional well-being, with a transient poverty component of 25.8 per cent and
15.0 per cent respectively, ¬‚uctuate less than income poverty, with a transient
poverty component of 33.0 per cent. Also, the well-being of adults seems to be
much more stable than the well-being of children. Whether stable well-being
is positive or negative from a normative perspectives depends on whether an
individual is poor or non-poor in a certain well-being dimension. For the poor,
steady indicators mean poverty traps; for the non-poor steady indicators mean
higher permanent well-being. But for all human development indicators
(except education for adults) there is a signi¬cant transient component, i.e. it
is well worth studying the dynamics of non-income dimensions of well-being.
Some important cautionary remarks have to be made concerning the inter-
pretation of poverty dynamics across different well-being dimensions. The
¬rst issue relates to the fact that we often only have a two-wave panel
at hand, where high-income transient poverty might largely be caused by
general economic development. In our case, Vietnam experienced signi¬cant
economic growth which led to a massive decrease in income poverty there
between 1992/3 and 1997/8 (see also Bonschab and Klump, 2007), with the
headcount poverty rate falling from about 61 per cent to 34 per cent. All
measures of non-income poverty show much smaller improvements.
11
We thus do not apply equivalence scales. White and Masset (2003) have recently shown
the ˜bias™ induced by ignoring household size and composition on poverty pro¬les for
Vietnam in a static context. In a further study it might hence be interesting to analyse the
impact of equivalence scales (and hence also household dynamics) on measured income (or
consumption) poverty dynamics; see also the discussion on household size below.




88
Measuring Chronic Non-Income Poverty

Table 4.1. Poverty rates and dynamics

Income Nutrition Education
Total Adult Child Total Adult Child Total Adult Child

Poverty 1992 61.5 56.6 69.7 43.2 33.6 54.9 58.1 64.1 29.4
Poverty 1997 34.2 31.2 40.3 34.5 30.9 41.5 49.7 57.9 17.7
Chronic 31.4 27.8 38.2 26.0 23.0 32.6 43.7 57.9 14.6
Transient 33.0 32.2 33.6 25.8 18.6 29.7 15.0 6.2 32.9
Non-poor 35.6 40.0 28.2 48.2 58.4 37.7 31.3 35.9 52.5



With a two-wave panel in this economic boom environment, it is therefore
dif¬cult to distinguish whether high-income poverty dynamics are caused by
income ¬‚uctuations or by a move out of structural poverty of a large part of
the population. Likewise, we do not know whether we observe higher chronic
non-income poverty because human development indicators are more stable
(i.e. are less volatile) or because they adjust more slowly (i.e. with some delay)
than income indicators to economic development. The interesting question
here then is whether dynamics of non-income indicators rather re¬‚ect past,
whereas income poverty dynamics re¬‚ect current, poverty dynamics.
In addition, differences between income and non-income poverty dynamics
might also be explained by the somewhat ˜arbitrarily set™ level, i.e. poverty
line, in different poverty dimensions. In other words, chronic poverty rates are
certainly positively correlated with the extent of total poverty and negatively
related to poverty reduction (or increases) over time: i.e. the higher the static
poverty rate the higher chronic poverty, and the higher poverty reduction
(or increase) the higher transient poverty. Hence, differences in the extent
of chronic poverty using income and non-income indicators might just stem
from the fact that the extent of total (static) poverty rates is different.
We deal with this potential measurement problem by equalizing poverty
rates across the different indicators. For example, Table 4.2 shows ˜¬tted™
income poverty rates for the case of nutritional and income poverty, where
we ¬rst align income poverty rates to the level of nutrition poverty in 1992,
i.e. the consumption poverty line is endogenously set so that total static
income poverty rate is equal to the level of static nutrition poverty in the ¬rst
year. If we do this, the share of transitory income poverty remains about the
same and much higher than the share of transitory nutrition poverty, while
the share of chronically poor falls as expected. Thus the higher transitory
component is not related to the initial setting of the poverty line. If, however,
we equalize income total (static) poverty rates to nutritional (static) poverty
rates in both years 1992 and 1997, the differences between income and non-
income chronic and transitory poor largely disappears; thus much of the
transitory component of income poverty is indeed related to a quicker escape
from income than non-income poverty in Vietnam during that era. However,



89
Isabel Günther and Stephan Klasen

Table 4.2. Poverty dynamics using ¬tted income poverty rates

Income Nutrition Education Income adj. to Income adj. to
Nutrition Education

Poor 1992 61.5 43.2 58.1 43.2 43.2 58.1 58.1
Poor 1997 34.2 34.5 49.7 34.5 17.0 30.8 49.7
Chronic poor 31.4 26.0 43.7 26.4 14.6 27.7 41.5
Transient poor 33.0 25.8 15.0 25.1 31.2 33.6 24.9
Non-poor 35.6 48.2 31.3 48.5 54.2 38.7 33.6

Note: In the ¬rst set of adjusted poverty rates we adjust the income poverty rate to the nutritional (educational)
poverty rate in the ¬rst year and then in¬‚ate it with the in¬‚ation rate implied by the of¬cial poverty line in¬‚ation
between 1992 and 1997, while in the second adjustment we adjust income poverty rates in both years to
nutritional (educational) poverty.


if we adjust income poverty rates to education poverty rates in both years, still
the transient component of income poverty is much higher than the transient
component of educational poverty, indicating that educational well-being is
indeed much more stable over time than income poverty (and nutritional
poverty).
Two further measurement issues that might explain the higher transient
component in income poverty dynamics are household dynamics and mea-
surement error. As stated above, we consider the total household for a calcu-
lation of per capita incomes and thus income poverty, while we only consider
individuals present in both surveys for our non-income analysis. Household
dynamics, i.e. increasing or decreasing household size, will have a signi¬cant
in¬‚uence on per capita income and thus affect poverty dynamics (by affect-
ing the denominator by which existing household income is divided or by
additionally affecting the numerator if the additional person is contributing
incomes), while they do not directly affect the non-income well-being of
individuals tracked (see discussion in section 4.3). With regard to measure-
ment error, income (or consumption) is likely to be measured with higher
measurement error than non-income indicators; thus a considerable part of
transient income poverty might indeed be caused by measurement error. And
with only a two-wave panel at hand there is little scope for appropriate instru-
ments to control for measurement error (see Woolard and Klasen, 2005, for
a discussion). Bhatta and Sharma (2006) have nevertheless lately applied the
proposed method of Luttmer (2002), of error-adjusted consumption measures,
to a two-wave panel in Nepal, which might deserve further consideration,
although some rather stringent assumptions have to be made.



4.5.2. Correlation of poverty dynamics
Even if national levels of income and non-income poverty were the same at a
point in time or across time, it could still be the case that the income chronic



90
Measuring Chronic Non-Income Poverty

Table 4.3. Correlation of income and non-income dynamics

Income Nutrition Education
Chronic Transient Non-poor Chronic Transient Non-poor

Chronic 33.5 27.5 39.0 49.8 18.7 31.6
Transient 26.9 27.5 45.6 43.4 15.8 40.9
Non-poor 18.5 22.6 58.9 39.3 11.3 49.3




(transient) poor are different from the non-income chronic (transient) poor,
i.e. depending on the measures used we might identify different households
(individuals) as chronically poor. This is most important from a policy per-
spective as it would affect the targeting of anti-poverty policies. 12
Table 4.3 illustrates the correlation between income and the diverse non-
income poverty dynamics. The numbers show row percentages: that is, they
show the percentage of the income chronic (transient, non-) poor that are
also non-income chronic (transient, non-) poor, i.e. each row sums to 100 per
cent. 13
Although there is a positive correlation between income and non-income
poverty dynamics the correlation is quite low. In fact, it is astounding how
many chronic income poor are never poor in a nutrition and education
perspective and vice versa. 14 For example, 39.0 per cent of the chronic income
poor are never nutritionally poor. The correlation is even lower for transient
poverty. For example, the likelihood to be nutritionally transient poor does
not (or not much) increase if the individual is income transient poor: 27.5 per
cent of the chronic income poor as well as only 27.5 per cent of the transient
income poor are also nutritionally transient poor.
One could again argue that part of the low correlation between income and
non-income indicators is a consequence of general differences in poverty lev-
els (see previous section). However, if we use ¬tted income poverty dynamics,
i.e. we set income poverty rates in 1992 and 1997 equal to nutritional and edu-
cational poverty, the correlation between income and non-income poverty
dynamics does not improve signi¬cantly. This low correlation between the
income and non-income poverty dynamics even if we use ¬tted income
poverty rates could then be explained by two other major factors which we
explore in turn: either there is already a low correlation between different
dimensions of static poverty (see Table 4.4) or different dimensions of well-
being show very different dynamics (see Table 4.5).

12
See Klasen (2000) for a discussion in a static context.
13
Alternatively, one could have calculated the percentage of the non-income chronic
(transient, non-) poor which are also income chronic (transient, non-) poor. As we came
to the same conclusions applying this latter approach, we only report the former.
14
See Baulch and Masset (2003) for a similar ¬nding.




91
Isabel Günther and Stephan Klasen

Table 4.4. Correlation of static income and non-income poverty

Nutrition 1992 Education 1992
Poor Non-poor Poor Non-poor

Income 1992
Poor 29.9 31.6 31.3 27.9
Non-poor 13.3 25.3 18.7 22.1
Income 1997
Poor 14.5 19.9 17.0 16.8
Non-poor 20.1 45.7 26.7 39.5




Table 4.4 shows the static correlation between income and non-income
poverty in 1992 and 1997. Each year and each human development dimen-
sion sums to 100 per cent. It can be observed that the income poor are not
necessarily the non-income poor. For example, in 1992 29.9 per cent of the
population is both income and nutrition poor whereas 25.3 per cent of the
population is neither income nor nutrition poor. However, 44.9 per cent of
the population is either income poor and not nutrition poor or nutrition
poor but not income poor. In 1997, due to signi¬cant economic develop-
ment in Vietnam in the 1990s, the share of the poor in both dimensions
has decreased whereas the share of the non-poor in both dimensions has
signi¬cantly increased, but still 40.0 per cent of the population is only poor
in one dimension but not poor in the other. The same trends can be observed
if we analyse educational poverty instead. Thus the extent of differences in
static poverty is very large, in fact larger than in some other studies where
income poverty was compared with composite non-income measures of well-
being (e.g. Klasen, 2000). 15
To separate differences in static poverty from differences in dynamics across
the various well-being dimensions, in Table 4.5 we analyse the correlation
of different poverty dynamics of only those individuals who show the same
static well-being in 1992. More precisely, we only analyse those individuals
who were either both income and non-income poor or neither income nor
non-income poor in 1992. Hence we exclude those individuals who were
poor in one but not in the other well-being dimension. The ¬gures show row
percentages, i.e. show the percentage of income chronic (transient, non-) poor
that are non-income chronic (transient, non-) poor. It should be clear that if
we exclude individuals who were initially income poor but not non-income


15
Part of this difference is, as discussed above, surely related to the ˜noise™ in the anthro-
pometric indicator which only gives a probabilistic assessment of a true nutritional de¬cit of
an individual.




92
Measuring Chronic Non-Income Poverty

Table 4.5. Correlation of income and non-income dynamics

Income Nutrition Education
Chronic Transient Non-poor Chronic Transient Non-poor

Chronic poor 65.3 34.7 0.0 87.9 12.1 0.0
Transient 52.9 38.2 8.9 80.9 15.0 4.1
Non-poor 0.0 11.4 88.6 0.0 12.1 87.9

Note: Only initial poor/non-poor in both income and non-income dimension are considered.


poor (and vice versa), there can be no individuals who are chronically poor in
one dimension but non-poor in another dimension.
If we analyse differences in pure poverty dynamics, i.e. poverty dynamics
controlled for differences in static poverty, the correlation between poverty
dynamics of income and non-income indicators increases signi¬cantly. Espe-
cially the income non-poor also seem to stay non-income non-poor: approxi-
mately 80 per cent of the income non-poor also stay non-poor in other dimen-
sions of well-being. To a large extent also the chronic income poor remain
(chronically) poor in non-income dimensions. In contrast the transient
income poor, i.e. mostly those individuals who move out of poverty, often
stay chronically poor in other dimensions, which could be caused by delayed
dynamics, where non-income indicators change after income well-being has
changed (i.e. transient non-income poverty re¬‚ects past poverty dynamics
whereas transient income poverty re¬‚ects current poverty dynamics).
In general, though, the dynamics of income and non-income poverty
are more similar than their static correlation, which is an interesting and
important ¬nding. It suggests that the (unmeasured) characteristics that affect
this lack of static correlation between income and non-income poverty do
not change much over time as the dynamic correlation for those who were
identi¬ed as poor/non-poor in both dimensions is quite similar.


4.6. Intra-Household Poverty Dynamics

As discussed above, a particular advantage of examining non-income dimen-
sions of well-being is the ability to study intra-household differences in well-
being levels and trends. In this section we explore household non-income
poverty dynamics, i.e. analyse the difference between (aggregate) household
and individual non-income poverty dynamics, which cannot be captured by
income or consumption measures of poverty dynamics which always assume
that either everyone or no one in the household is income poor. Differences
in household and individual poverty dynamics might, as already discussed
above, also be partly responsible for the very low correlation between income




93
Isabel Günther and Stephan Klasen

Table 4.6. Household non-income poverty dynamics

Nutrition Education
Total Adult Child Total Adult Child

Homogeneous non-income dynamics
Chronic poor 2.5 7.1 13.7 9.1 37.7 2.2
Transient poor 1.4 4.5 9.5 0.1 1.1 10.3
Non-poor 11.7 34.5 20.7 14.8 18.0 53.1
15.6 46.1 43.9 24.0 56.8 65.6

Heterogeneous non-income dynamics
Transient and chronic poor 6.8 6.9 18.8 10.9 5.8 2.7
Transient and non-poor 22.4 18.9 16.2 8.2 3.4 28.1
Chronic and non-poor 55.3 28.1 21.1 56.8 34.0 3.7
84.5 53.9 56.1 76.0 43.2 34.4




and non-income poverty dynamics, with the former measuring household
and the latter measuring individual poverty dynamics.
Table 4.6 shows intra-household poverty dynamics of the various non-
income indicators. The indicated percentages refer to individuals (or adults
and children) who live in households where all members are chronically,
transient, or non-poor (homogeneous poverty dynamics) or where some are
transient while others are chronically poor or non-poor, or where some house-
hold members are chronically poor whereas others are non-poor (heteroge-
neous poverty dynamics). One should be very cautious looking at the total
population and should rather analyse adults and children separately, as a lot
of differences in poverty dynamics between adults and children are caused by
differences in measurement (e.g. the nutritional status of adults is measured
weight over height, whereas the nutritional status of children is measured
height over age).
Whether we look at nutrition or education, the percentage of individuals
who live in households where all adult or child members show the same
poverty dynamics is only around 40“60 per cent. What is most surprising
is that up to one-third of individuals even live in households where some
household members are never poor in a particular non-income dimension
while others are always poor in that same dimension.
This high heterogeneity of individual poverty dynamics within house-
holds can also explain part of the low correlation of income and non-
income poverty dynamics at the micro-level (section 4.4.2) as well as on the
aggregate macro-level (section 4.4.1). In contrast to non-income indicators,
income indicators ignore differences in poverty dynamics within households.
We illustrate this in Table 4.7, where we compare individual nutritional
(and educational) poverty rates with per capita household average nutritional
(and educational) poverty rates. If we use the household average instead



94
Measuring Chronic Non-Income Poverty

Table 4.7. Average household poverty dynamics

Nutrition Education
Individuala Householdb Individuala Householdb

Poverty 1992 33.6 25.7 64.1 77.6
Poverty 1997 30.9 21.6 57.9 72.6
Chronic 23.0 14.5 57.9 70.4
Transient 18.6 18.9 6.2 9.2
Non-poor 58.4 66.6 35.9 20.4

a
Poverty rates refer to the individual BMI and years of schooling for adults 18+.
b
Poverty rates refer to per capita average household BMI and schooling. Rates only for adults of age 18+.



of individual rates the transient poverty rate (relative to the chronic part)
becomes signi¬cantly larger. So part of the lower transient non-income
poverty rate”in comparison to income poverty”stems from the fact that
we use individual instead of average household (scaled up to household
members) well-being indicators. If one individual improves his or her welfare
all other household members become better/worse off as well, so we arti¬cially
increase transient poverty if we work with household means. Also absolute
chronic poverty rates change signi¬cantly if we work with household averages
instead of individual poverty rates. But whereas for the nutritional poverty
rate the chronic component decreases”in comparison to individual rates”
for the educational poverty rate chronic poverty would signi¬cantly increase
if we worked with household poverty rates.
Lastly, non-income well-being indicators, or intra-household poverty
dynamics, can also be used to analyse long-term (intergenerational) poverty
dynamics, which is usually not possible with income indicators. Intergenera-
tional chronic poverty, which refers to poverty that is passed from one gener-
ation to the next, i.e. the most severe form of chronic poverty, can be assessed
by comparing the well-being of two generations within the same households.
Table 4.8 shows nutritional and educational poverty for all households, where
at least two generations were present in the household. 16 Poor elderly indicates
the poverty rate of individuals of the older generation within a household,
whereas Poor young refers to the poverty rate of individuals belonging to
the younger generation within a household. Poor refers to individuals that
are living in households where both the older and younger generations are
poor, i.e. where intergenerational chronic poverty persists. By de¬nition, all
generations within the same household are either income poor or not. How-
ever, there is quite a signi¬cant share of individuals who live in households
where one generation is non-income poor whereas the other generation is

16
Poverty rates were calculated based on the average consumption, BMI, and educational
level of adults older than 18 years belonging to one of the two generations within households.




95
Isabel Günther and Stephan Klasen

Table 4.8. Intergenerational chronic poverty (1997)

Income Nutrition Education

Poor elderly 31.1 24.4 67.6
Poor young 31.1 22.6 55.5
Poor 31.1 8.9 50.7
Poor/non-poor 0.0 34.8 31.9
Non-poor 69.9 56.4 17.4

Note: Rates are shown for the cross-section data of 1997. However,
we obtain the same trends if we use the data from 1992 instead.


not non-income poor. But a large part of individuals live in households
where particularly educational poverty is passed from one generation to the
next, and we should very much be concerned about these households where
poverty persists over very long time horizons.



4.7. Multidimensional Poverty as Chronic Poverty

Several authors have argued that chronic poverty might also be characterized
by multidimensional poverty (see e.g. Hulme, Moore, and Shepherd, 2001),
i.e. individuals who are poor in several dimensions are more likely to stay
chronically poor. We test this hypothesis by analysing the correlation of the
number of dimensions an individual is deprived of in 1992 and 1997 in
Table 4.9. For example 7.3 per cent of the total population have been poor in
one well-being dimension in 1992 and in one well-being dimension in 1997,
whereas 4.1 per cent of the population have been poor in all three dimensions
(income, nutrition, and education) in 1992 and 1997.
What is striking is that although the correlation of poverty dynamics of dif-
ferent well-being indicators seems to be rather low, i.e. moving out of income
poverty does not mean moving out of non-income poverty (and vice versa),
the number of well-being dimensions an individual is deprived of seems to
be quite stable over time. Fifty per cent of individuals have not changed the
number of dimensions in which they are poor (the sum of the diagonal shares)

Table 4.9. Chronic poverty as multidimensional poverty

1992 1997
Non- One Two Three
dimensional dimensional dimensional dimensional

Non-dimensional 11.6 2.6 0.4 0.0
One dimensional 10.5 19.4 4.0 0.5
Two dimensional 3.9 14.7 15.8 3.5
Three dimensional 0.4 2.9 6.2 4.1




96
Measuring Chronic Non-Income Poverty

and very few (8.2 per cent) have changed by more than one dimension. This
¬nding could to some extent even explain the low correlation of income and
non-income poverty dynamics, if we assume that poor individuals alternate
between, for example, low educational or low health functionings or between
income and non-income poverty. So here the extent of poverty would be
de¬ned as the number of well-being dimensions a person is deprived of. If
she is poor in one dimension in one year it is very likely that she is also
poor in one dimension in the following years, but the dimension can change.
One intriguing interpretation would be that individuals are forced to choose
between different forms of deprivation and make different choices over time.
This issue certainly deserves closer attention in future research.


4.8. Conclusion and Further Research

The main ¬ndings from this exploratory analysis to study chronic poverty
and/or poverty dynamics from a non-income perspective are, ¬rst, that there
are sound theoretical as well as practical empirical arguments for moving in
such a direction. It generates important new insights about the dynamics
of well-being outcomes over time, their relationship to incomes, and intra-
household and intergenerational dynamics. In particular, in our empirical
assessment there is more dynamics in non-income dimensions of poverty
than commonly presumed, although non-income poverty is certainly more
stable over time than income poverty. Moreover, the correlation between
chronic poverty in the income and non-income dimensions is very low. This
seems to be mostly caused by the low static correlation between the two rather
than by different dynamics over the observed period. Fourth, we observed a
rather high heterogeneity in intra-household non-income poverty dynamics,
which would not be captured by income poverty measures. Last, the number
of well-being dimensions individuals are deprived of is surprisingly stable
over time. Given the limitations of the data we had at our disposal, these
are interesting ¬ndings worth exploring more.
But clearly one implication of our research is that more effort must be
directed into generated comparable panel datasets that fully capture impor-
tant non-income well-being outcomes in a comparable fashion. Among the
most important improvements to tackle are better measures of health status
(see Schultz, 2002, 2003, for possible suggestions) and the inclusion of educa-
tional test scores for all in the household.
In addition, this largely descriptive analysis leaves a number of questions
unanswered which should be the topic for further research. Most important
is a formal regression-based analysis of the determinants of income and
non-income poverty dynamics to further understand the surprisingly low
correlation between the two as well as the high heterogeneity of poverty



97
Isabel Günther and Stephan Klasen

dynamics within households. To date, most related regressions have only
examined the determinants of chronic and transient income poverty where
some non-income dimensions of well-being (particularly human assets such
as health and education) are seen as important determinants (e.g. Woolard
and Klasen, 2005). Such regression approaches could be extended to also
explain dynamics of non-income poverty. Controlling for measurement error
and endogeneity will clearly be an issue here, which can be more easily
achieved if one can use lagged values as instruments in panels that have more
than two waves. Such analyses should usefully consider the actual levels of
income and non-income deprivation rather than be based on dichotomous
poverty de¬nitions as used here in our exploratory analysis.
Moreover, one can more systematically examine whether some households
are chronically worse at turning incomes into non-income achievements. This
can be done by examining the persistence of positive and negative residuals of
non-income regressions among households across time or applying quantile
regressions. This would uncover and de¬ne households that are chronically
underperforming in turning incomes into functionings as chronically poor.
Secondly, the question of household structure dynamics and equivalence
scales deserves closer examination. As shown for example by Woolard and
Klasen (2005), changes in household size and structure are an important
determinant of income mobility over time and we also know that static
poverty assessments are sensitive to equivalence scale assumptions. Both of
these issues were raised here but deserve further analysis, particularly when
comparing income to non-income poverty dynamics.
Thirdly, one can examine the whole distribution of income and non-income
well-being dynamics, i.e. using continuous measures rather than dichotomous
indicators to study chronic poverty in a non-income dimension. Here the
research of Grosse, Harttgen, and Klasen (e.g. Grosse, Harttgen, and Klasen,
2005; Klasen, 2005) in combination with the work of Grimm (2006) could be
extended to study non-income poverty dynamics across the entire well-being
distribution of households.
A last interesting extension of our work would be to derive multidimen-
sional measures of non-income poverty dynamics, which go beyond a partial
ordering of well-being outcomes. The challenging question here is not only
how to weight and aggregate different well-being dimensions but in addition
how to weight and aggregate different time dimensions. Such work could
build on studies by Bossert and D™Ambrosio (2006) and Chakravarty and
D™Ambrosio (2006) who axiomatically derive relative and absolute measures
of social exclusion, i.e. chronic capability failure. For them, social exclusion
is the (weighted) sum of individual functionings from which an individual
is excluded over time. The chapter is much concerned with the aggregation
to a social exclusion score for the society and with comparisons with other
societies using dominance relations, which could be a helpful start for such



98
Measuring Chronic Non-Income Poverty

work. In this context, it might also be fruitful to combine the study of Duclos,
Sahn, and Younger (2006) with the one of Gr¤b and Grimm (2006), with the
former concentrating on robust multidimensional and the latter focusing on
robust multi-period poverty comparisons in non-income dimensions.




References

Atkinson, A. B., and Bourguignon, F. (2000), Handbook of Income Distribution,
Amsterdam: North Holland.
Basu, K., and Foster, J. (1998), ˜On Measuring Literacy™, Economic Journal, 108: 1733“49.
Baulch, B., and Hoddinott, J. (2000), ˜Economic Mobility and Poverty Dynamics in
Developing Countries™, Journal of Development Studies, 36(6): 1“24.
and Masset, E. (2003), ˜Do Income and Non-Income Indicators Tell the Same Story
about Chronic Poverty? A Study of Vietnam in the 1990s™, World Development, 31(3):
441“53.
Bhatta, S. D., and Sharma, S. K. (2006), ˜The Determinants and Consequences of
Chronic and Transient Poverty in Nepal™, CPRC Working Paper 66, Manchester:
Chronic Poverty Research Center.
Bonschab, T., and Klump, R. (2007), ˜Pro-Poor Growth in Vietnam: Explaining the
Spatial Differences™, in M. Grimm, A. McKay, and S. Klasen (eds.), Determinants of
Pro-Poor Growth: Determinants of Pro-Poor Growth: Analytical Issues and Findings from
Country Cases, London: Palgrave Macmillan.
Bossert, W., and D™Ambrosio, C. (2006), ˜Deprivation and Social Exclusion™, Economica,
forthcoming.
Chakravarty, S., and D™Ambrosio, C. (2006), ˜The Measurement of Social Exclusion™,
Review of Income and Wealth, 52(3): 377“98.
Clark, D., and Hulme, D. (2005), ˜Towards a Uni¬ed Framework for Understanding the
Depth, Breadth and Duration of Poverty™, Global Poverty Research Group Working
Paper 20, Manchester: Chronic Poverty Research Center.
Deaton, A. (1997), The Analysis of Household Surveys, Baltimore: Johns Hopkins
University Press.
and Zaidi, S. (2002), ˜Guidelines for Contructing Consumption Aggregates for
Welfare Analysis™, LSMS Working Paper No. 135, Washington: World Bank.
Duclos, J. Y., Sahn, D., and Younger, S. D. (2006), ˜Robust Multidimensional Spatial
Poverty Comparisons in Ghana, Madagascar, and Uganda™, World Bank Economic
Review, 20(1): 91“113.
Friedman, M. (1957), A Theory of the Consumption Function, Princeton: Princeton
University Press.
Gibson, J. (2001), ˜Measuring Chronic Poverty without a Panel™, Journal of Development
Economics, 65: 244“63.
Gr¤b, J., and Grimm, M. (2006), ˜Robust Multiperiod Poverty Comparisons™, paper
prepared for the Chronic Poverty Research Center.
Grimm, M. (2007), ˜Removing the Anonymity Axiom in Assessing Pro-Poor Growth™,
Journal of Economic Inequality, 5(2): 179“97.



99
Isabel Günther and Stephan Klasen

Grosse, M., Harttgen, K., and Klasen, S. (2005), ˜Measuring Pro-Poor Growth Using Non-
Income Indicators™, Ibero-America Institute Working Paper 132, Göttingen: University
of Göttingen.
(2008), ˜Measuring Pro-Poor Growth in Non-Income Dimensions™, World
Development, 36(6): 1021“47.
Haddad, L., and Kanbur, R. (1990), ˜How Serious is the Neglect of Intra-Household
Inequality?™, Economic Journal, 100: 866“81.
Henderson, R. M. (2005), ˜The Bigger the Healthier: Are the Limits to BMI Risk Changing
over Time?, Economics and Human Biology, 3: 339“66.
Hulme, D. (2003), ˜Chronic Poverty and Development Policy: An Introduction™, World
Development, 31(3): 399“402.
and McKay, A. (2005), ˜Identifying and Understanding Chronic Poverty: Beyond
Income Measures™, Manchester: Chronic Poverty Research Center.
Moore, K., and Shepherd, A. (2001), ˜Chronic Poverty: Meanings and Analytical
Frameworks™, CPRC Working Paper 2, Manchester: Institute of Development Policy
and Management.
and Shepherd, A. (2003), ˜Conceptualizing Chronic Poverty™, World Development,
31(3): 403“23.
Jalan, J., and Ravallion, M. (1996), ˜Transient Poverty in Rural China™, Policy Research
Working Paper 1616, Washington: World Bank.
Jensen, R. (2000), ˜Agricultural Volatility and Investments in Children™, American Eco-
nomic Review, 90(2): 399“404.
Klasen, S. (2000), ˜Measuring Poverty and Deprivation in South Africa™, Review of Income
and Wealth, 46: 33“58.
(2005), ˜Economic Growth and Poverty Reduction: Measurement and Policy Issues™,
OECD Development Centre Working Paper 246, Paris: OECD.
(2007a), ˜Gender-Related Indicators of Well-Being™, in M. McGillivray (ed.), Human
Well-Being: Concept and Measurement, London: Palgrave, 167“92.
(2007b), ˜Poverty, Undernutrition, and Child Mortality: Some Inter-regional
Puzzles and their Implications for Research and Policy™, Journal of Economic Inequality,
forthcoming.
(2008a), ˜Economic Growth and Poverty Reduction: Measurement Issues using
Income and Non-Income Indicators,™ World Development, 36(3): 420“45.
(2008b), ˜Poverty, Undernutrition and Child Mortality: Some Inter-regional Puzzles
and their Implications for Research and Policy™. Journal of Economic Inequality, 6(1):
89“115.
Luttmer, E. F. P. (2002), ˜Measuring Poverty Dynamics and Inequality in Transition
Economies: Disentangling Real Events from Noisy Data™, NBER Working Paper.
McKay, A., and Lawson, D. (2003), ˜Assessing the Extent and Nature of Chronic Poverty
in Low Income Countries: Issues and Evidence™, World Development, 31(3): 425“39.
Moradi, A. (2006), ˜Nutritional Status and Economic Development in Sub-Saharan
Africa, 1950“1980™, working paper, Oxford: University of Oxford.
OECD (2000), Literacy in the Information Age, Paris: OECD.
Ramos, X., and Silber, J. (2005), ˜On the Application of Ef¬ciency Analysis to the Study
of the Dimensions of Human Development™, Review of Income and Wealth, 51: 285“
310.




100
Measuring Chronic Non-Income Poverty

Schultz, T. P. (2002), ˜Health and Labor Force Participation of the Elderly in Taiwan™,
Yale University Economic Growth Center Discussion Paper No. 846.
(2003), ˜Human Capital, Schooling, and Health™, Economics and Human Biology, 1:
207“22.
Sen, A. (1985), The Standard of Living, Cambridge: Cambridge University Press.
(1999), Development as Freedom, Oxford: Oxford University Press.
Townsend, R. (1995), ˜Consumption Insurance: An Evaluation of Risk-Bearing in Low-
Income Economies™, Journal of Economic Perspectives, 9(3): 83“102.
UNICEF (1998), The State of World™s Children: Focus on Nutrition, New York: UNICEF.
White, H., and Masset, E. (2003), ˜Constructing the Poverty Pro¬le: An Illustration of the
Importance of Allowing for Household Size and Composition in the Case of Vietnam™,
Development and Change, 34(1): 105“26.
WHO (1995), Physical Status: The Use and Interpretation of Anthropometry, WHO Technical
Report Series No. 854, Geneva: WHO.
Woolard, I., and Klasen, S. (2005), ˜Income Mobility and Household Poverty Dynamics
in South Africa™, Journal of Development Studies, 41: 865“97.
Yaqub, S. (2003), ˜Chronic Poverty: Scrutinizing Patterns, Correlates and Explorations™,
CPRC Working Paper 21, Manchester: University of Manchester.
Zeller, M., Sharma, M., Henry, C., and Lapenu, C. (2006), ˜An Operational Method for
Assessing the Poverty Outreach Performance of Development Policies and Projects™,
World Development, 34: 446“64.




101
5
The Construction of an Asset Index
Measuring Asset Accumulation in Ecuador

Caroline Moser and Andrew Felton




5.1. Introduction

In the past decade development economists have increasingly advocated
the use of assets to complement income and consumption-based measures
of welfare and wealth in developing countries (Carter and May, 2001;
Filmer and Pritchett, 2001). Income has long been the favoured unit of
welfare analysis, because it is a cardinal variable that is directly compara-
ble among observations, making it straightforward to interpret and use in
quantitative analysis. However, by the 1990s this was often superseded by
consumption-based measures (Ravallion, 1992). The analysis of assets and
their accumulation is intended to complement such measures, by extending
our understanding of the multidimensional character of poverty and the
complexity of the processes underlying poverty reduction (Adato, Carter, and
May, 2006).
Closely linked to the asset-based approach is recent methodological work
on the measurement of assets with a range of new techniques developed to
capture aggregate ownership of different assets into a single variable. The
objective of this ˜technical™ chapter is to contribute to the debate about the
measurement of assets. It describes the particular methodology developed
to construct an asset index based on a longitudinal panel dataset from



The authors gratefully acknowledge Michael Carter, who as adviser to the Guayaquil
project has provided invaluable guidance and positive feedback as we have grappled with
creating the asset index. Thanks also to James Pickett, John Hoddinott, Jesko Hentschel,
Michael Woolcock, and Peter Sollis for advice. The latest stage of the Guayaquil project has
been supported by the Ford Foundation, New York. Particular thanks to vice president Pablo
Farias for his commitment to this work.




102
Asset Accumulation in Ecuador

Guayaquil, Ecuador. It then outlines its application in terms of the different
components of the asset index, before concluding by identifying several con-
tinuing methodological problems.



5.2. Contextual Background

5.2.1. The research methodology
The construction of an asset index is grounded in a research project on
˜Intergenerational asset accumulation and poverty reduction in Guayaquil,
Ecuador between 1978 and 2004™. A community study such as this, which
combines a range of qualitative, participatory, and quantitative method-
ological approaches used over a twenty-six-year research period, poses chal-
lenges relating to its statistical robustness, or representativeness. This was
also the case in an earlier research phase when the data were included in
a World Bank study on the ˜social impact™ of structural adjustment reforms
in four poor urban communities in different regions of the world that
included not only Guayaquil, Ecuador, but also Lusaka, Zambia, Budapest,
Hungary, and Metro Manila, the Philippines (Moser, 1996, 1998). At the
time the results were dismissed by World Bank economists as not represen-
tative at the national level, nor robust in terms of cross-country compar-
isons; at best they provided interesting case study ˜anecdotal information™
on community and household coping strategies in ˜crisis situations™ (Moser,
2002).
To address this challenge the research methodology for this ¬nal study
builds on earlier cross-disciplinary combined methodologies including the
pioneering work of Ravi Kanbur to address the ˜qual“quant™ divide (Kanbur,
2002) but pushes the envelope further by including the econometric mea-
surement of the quantitative data on capital assets. The methodology, which
we have called ˜narrative econometrics™, combines the econometric measure-
ments of change with in-depth anthropological narratives that identify the
social relations within households, communities, and broader institutional
structures that in¬‚uence well-being and assist in identifying the associated
causality underpinning economic mobility. In so doing we also seek to
develop methodological tools that can bridge the divide in current debates
about the limitations of measurement-based poverty analysis that disregards
context and therefore ˜cannot address the dynamic, structural and relational
factors that give rise to poverty™ (Harriss, 2007; see also Green and Hulme,
2005; Green, 2006). This chapter, however, is limited to the elaboration of the
index methodology and therefore complements further analysis that seeks to
bring together the ˜econometric™ and the ˜narrative™ (see for instance Moser
and Felton, 2007).



103
Caroline Moser and Andrew Felton

5.2.2. Assets and income
While economists often use income to measure wealth, welfare, and other
indicators of well-being, income data has limitations in both accuracy
and measurement, particularly in the context of developing countries. For
instance, for people living in informal labour markets incomes are often
highly variable. Income can be seasonal, such as when earned from farming
or the tourist market, or just variable and lumpy for small-business owners.
Taking a snapshot of income at one point in time may therefore produce a
less reliable picture of these types of workers than those who receive regular
pay cheques. Furthermore, they may be engaged in barter and other non-
monetary forms of trade. In all of these cases there is a high potential for
error in data based on the recollection and value of all sources of income. This
means that income itself does not necessarily provide a reliable measure of
well-being.
Expenditures and consumption are also commonly used to measure well-
being (Chen and Ravallion, 2000; Ellis, 2000). Expenditures solve some of the
problems of income, such as seasonality. Households can save their income
from ¬‚ush times as a buffer against bad times. This ˜consumption smoothing™
is both theoretically appealing and has empirical regularity. Households also
tend to be more forthcoming about expenditures, which lack the sensitivity
that some have towards divulging income data. However, a number of the
same dif¬culties of income also apply to expenditure, such as measuring the
value of bartered goods. Work done for oneself, such as house improvement,
also tends to be missing from expenditures. In addition, although econo-
mists have shown that consumption data provide more robust information
on well-being than income data (particularly in rural areas), income data
are still used in a number of research studies such as in the Guayaquil
study. 1
When asking people what they own from a list of assets, there is often
less likelihood of recall or measurement problems. Furthermore, assets may
provide a better picture of long-term living standards than an income snap-
shot because they have been accumulated over time and last longer. However,
a list of assets lacks money™s advantages of cardinality and fungibility. The
following section explores the theoretical dif¬culties of creating a set of ˜asset™
variables.

1
Longitudinal anthropological research in this urban context revealed that even people™s
short-term recall of consumption expenditures was often inaccurate or underestimated. Peo-
ple buying many of their basic consumption items on a daily basis simply did not remember
what they spent. Data from expenditure diaries, for instance, proved to be widely inconsistent
with expenditure data from anthropological participant observation. In contrast, working in
a community where trust had been established, there was a high level of compatibility across
the 51 households in the panel data in terms of income relating to both formal and informal
sector earnings. For this reason the study used income measures.




104
Asset Accumulation in Ecuador

Suppose that a household™s capital portfolio can be measured in terms of a
number I of types of capital, C i , where i ∈ [1, 2, . . . I ]. Each type of capital C i
is composed of J types of assets ai,1 . . . ai,J . Each of these a™s may be measured
using a binary, ordinal, or cardinal variable. We want to assign a weight w to
each item and then sum up the weighted variables to arrive at our estimate of
C i , as in equation 5.1.

J
i, j i, j
wt an,t
i
Cn,t = (5.1)
j=1


where
n = Household number
i = Type of capital
j = Type of asset
t = Time period
The rest of this section describes different ways to measure the w™s.
Method 1: Prices. One intuitive way to weight the assets is to use monetary
i, j i, j i, j
values, so that wt = pt where pt is the price (or some other monetary
i, j i, j
J
j=1 wt an,t would then
measure of value) of asset (i,j) at time t. The sum
be the total monetary value of the household™s asset wealth. However, this
approach is problematic for some of the same reasons that income data are.
Price data can be dif¬cult to obtain in some contexts, especially in economies
that have high levels of barter. Even more fundamental is the problem that it
is dif¬cult or impossible to assign prices to intangible assets, such as human
or social capital. Of course, assigning any numbering to those types of capital
is tenuous, but the ordinal scale that we develop in this chapter seeks to
overcome the implied fungibility of prices.
Method 2: Unit values. Another method is to simply sum up the number of
assets owned, which is equivalent to setting w = 1 for each w. This method
has the virtue of simplicity, but also has the limitation of assigning equal
weight to ownership of each asset. For example, this method would assign
equivalent worth to owning a radio and a computer, although in reality their
contributions to the capital variable are surely different.
Method 3: Principle components analysis. Recently, development economists
have followed the recommendation made by Filmer and Pritchett (2001) to
use principle components analysis (PCA) to aggregate several binary asset
ownership variables into a single dimension. PCA is relatively easy to compute
and understand, and provides more accurate weights than simple summation.
The intuition underlying this method is that there is a latent (unobservable)
˜i
variable C for each type of capital C i that manifests itself through ownership
of the different assets ai,1 . . . ai,J . For example, suppose household n owns
˜i
asset ai,1 if C > wi,1 . It turns out that the maximum likelihood estimators



105
Caroline Moser and Andrew Felton

Regression minimizes dash lines
PCA minimizes gray lines




Figure 5.1. Difference between regression and PCA


of the w™s are the eigenvectors of the covariance matrix, also known as the
principle components of the dataset. 2 Usually only the eigenvector with the
highest eigenvalue is used, because it is the vector that provides the most
˜information™ about the variables. 3 The ¬rst eigenvector is the vector that
minimizes the squared distances from the observations to a line going through
the various dimensions.
This is an appealing method for combining variables for two reasons. First,
it is technically equivalent to a rotation of the dimensional axes, such that the
variance from the observations is minimized. This is equivalent to calculating
the line from which the orthogonal residuals are minimized. This is similar
to a regression in terms of minimizing residuals, but in this case the residuals
are measured against all of the variables, not just one ˜dependent™ variable.
Figure 5.1 demonstrates how regression minimizes the squared residuals from
a dependent variable to a line, while PCA minimizes the distances from points
in multidimensional space to a line.
The second reason that PCA is a valuable approach is that the coef-
¬cients have a fairly intuitive interpretation. The coef¬cient on any one
variable is related to how much information it provides about the other


2
In fact a correlation matrix is usually used to equally scale the variables and avoid
problems stemming from what measurement units are used.
3
It is also possible to use the sum of a number of eigenvectors, based on some criteria.
Using the sum of all the eigenvectors is equivalent to using unit coef¬cients for each variable.
Some statisticians recommend using all eigenvectors with eigenvalues greater than one;
others suggest the ˜scree test™. However, these are more complicated to interpret than using
just the ¬rst eigenvector (Jolliffe, 2002).




106
Asset Accumulation in Ecuador

variables. 4 If ownership of one type of asset is highly indicative of ownership
of other assets, then it receives a positive coef¬cient. If ownership of an
asset contains almost no information about what other assets the household
owns (its correlation coef¬cient is near zero), then it receives a coef¬cient
near zero. And if ownership of an asset indicates that a household is likely
to own few other assets, then it receives a negative coef¬cient. Higher and
lower coef¬cients mean that ownership of that asset conveys more or less
information about the other assets.
This makes PCA excellent for modelling a presumed underlying continuous
variable, such as wealth. If ownership of a certain asset is highly correlated
with owning the other assets that were asked about in the survey, then it is
probably correlated with owning other types of assets that were not in the
survey as well. To return to the earlier example, wealthy households are more
likely to own a computer than poor ones, but radio ownership is spread evenly
across the spectrum. Therefore, knowing that one household owns a computer
provides us with more information about that household™s wealth than a radio
does, and it receives a higher weighting.
Filmer and Pritchett (2001) fail to adequately address the important
methodological issue that the variables must positively correlate with the
latent variable, and with each other. If all the variables are positively corre-
lated, then the estimates will all be greater than or equal to 0 and bounded
at the top by the value of the ¬rst eigenvalue (which is itself less than or
equal to the number of variables in the matrix). If they are not, then the ¬rst
eigenvector may have negative values, which means that the estimated latent
variable would be reduced from ownership of an asset. This is only remedied
by interpreting ownership of those assets as a sign of lower wealth. If this
is plausible, then even negative values of estimated wealth are acceptable
because the estimated variable is ordinal and can either be used as is or
rescaled so that they are all positive.
Filmer and Pritchett (2001) use twenty-one types of assets from the Demo-
graphic and Health Surveys, covering both consumer durables and housing
stock, to create a single ˜wealth™ variable. They show that the resulting variable
has empirically plausible consequences and predicts school enrolment better
than expenditure. The robustness tests on asset indices conducted by Sahn
and Stifel (2003) demonstrate that the asset index reliably predicts poverty
and serves as a proxy for long-term wealth with less error than data on
expenditures.
Other papers advocate a variety of techniques. Sahn and Stifel (2003) use
factor analysis, which is designed more for data exploration than dimensional
reduction. Booysen, van der Berg, Burger, von Maltitz, and du Rand (2005)

4
This has a precise mathematical de¬nition in terms of Kullback“Leibler (1951) informa-
tion.




107
Caroline Moser and Andrew Felton

use multiple correspondence analysis (MCA), which they promote as better at
dealing with categorical variables than PCA. Finally, Kolenikov and Angeles
(2004) describe a new technique, polychoric principle components analysis,
which improves on regular PCA and is designed speci¬cally for categorical
variables. Unlike MCA it can also be used for continuous variables and is
especially appropriate for discrete data. It supposes that the discrete data are
observed values of an underlying continuous variable. Similar in spirit to
an ordered probit regression, polychoric PCA uses maximum likelihood to
calculate how that continuous variable would have to be split up in order
to produce the observed data.
Polychoric PCA has a number of advantages over regular PCA. For instance,
its coef¬cients are more accurately estimated than with regular PCA. 5 The
main advantage, however, comes from its use of ordinal data. Many assets
can be described as ordinal. Researchers often ask about the quality of con-
struction of a home, for example, which might be recorded on a 1“4 scale.
While Filmer and Pritchett (2001) advocate splitting this into four binary
variables, this introduces a large amount of distortion into the correlation
matrix, as the variables are automatically perfectly negatively correlated with
each other. Furthermore, the knowledge that the researcher brings”that some
values are better than others”is lost, as the PCA treats every variable as the
same. Polychoric PCA solves these problems by assigning each the value of
a discrete variable and ensuring that the coef¬cients of an ordinal variable
follow the order of its values.
Another advantage of polychoric PCA is that it allows us to compute coef-
¬cients of both owning and not owning an asset. This is desirable because
sometimes not owning something conveys more information than owning
it. If almost every household owns indoor plumbing except for the very
poorest, then the coef¬cient on owning indoor plumbing will be around zero
(since it does not help distinguish household wealth among those that own
it). However, not owning indoor plumbing will be negatively correlated to
ownership of other assets and the coef¬cient of not owning it will be highly
negative. This further distinguishes among wealth levels.



5.3. Multivariate Analysis

Most research so far has only used PCA and its related techniques to model
ownership of a single type of asset, usually a variant of ˜wealth™. However,
social scientists are often interested in examining portfolios that include
different asset types in order to better understand the speci¬c root causes

5
Kolenikov and Angeles (2004) run a Monte Carlo exercise on simulated data and ¬nd that
polychoric PCA predicts the ˜true™ coef¬cients more accurately than regular PCA.




108
Asset Accumulation in Ecuador

of poverty. Hulme and McKay (2005) provide an overview of techniques
used for multivariate asset analysis, brie¬‚y mentioning index construction
methods like PCA before moving on to a variety of other methods used by
economists, sociologists, and anthropologists. Most of the examples of multi-
variate asset analysis cited do not use PCA or other sophisticated techniques of
aggregating assets. For example, Klasen (2000) identi¬es fourteen components
of well-being and sums up the number that are unsatisfactory for a given
households to arrive at a ˜deprivation index™. However, as Hulme and McKay
point out,

while giving all components the same weight might appear to be ˜fair™, there is a
complex set of value judgments built into such an assumption. For example, can
nutrition (child stunting that may reduce an individual™s capabilities over her lifecourse)
be weighted the same as transport/mobility (where a low score may be a temporary
inconvenience)?

They identify similar issues with multidimensional frameworks by Clark and
Qizilbash (2002, 2005) and Barrientos (2003).
Among the papers that use PCA or similar methods, Sahn and Stifel (2003)
come closest to implementing a multidimensional approach. They categorize
their index components into three types of capital (household durables,
household characteristics, and human capital) but they then combine them
all together into a single index. Asselin (2002) also groups his variables into
categories (economy and infrastructure, education, health, and agriculture)
but then combines them all before his analysis. To the best of our knowledge,
no paper so far uses PCA on the components of each type of capital before
undertaking its analysis.



5.4. The Empirical Application of Polychoric PCA:
The Guayaquil Panel Dataset

Below we present an application of multivariate analysis on a panel dataset
of ¬fty-one households in a neighbourhood of Guayaquil, Ecuador, between
1978 and 2004. The analysis is based on the research presented in detail in
Moser and Felton (2007) on the measurement and analysis of four dimensions
of the ¬ve-dimensional asset framework (see for instance Carney, 1998; Moser,
1998; World Bank, 2000) in order to implement a quantitatively rigorous,
multidimensional approach to asset accumulation and poverty dynamics.
While the asset index is grounded in an extensive literature review on
livelihoods and asset accumulation (Moser, 2008), the speci¬c assets chosen
were determined by the questions available in the dataset. Here it is necessary
to recognize constraints relating to the fact that this was not originally de¬ned
as a study of asset accumulation when the ¬rst stage of ¬eld research was



109
Caroline Moser and Andrew Felton

undertaken living in the community. (This applies particularly to the social
capital variables and employment data.)


5.4.1. Data and contextual background
The data come from a research project that focused on household asset accu-
mulation strategies using twenty-six years of anthropological and sociological
research in a poor urban community, Indio Guayas, in Guayaquil, Ecuador.
Named after its community committee, this is an eleven-block neighbour-
hood area within the barrio of Cisne Dos, which in 2004 had an estimated
75,364 inhabitants (World Bank, 2004). Cisne Dos itself is one of a number
of working-class suburbs in the parroquia of Febres Cordero located on the
south-west edge of the city. This area, about seven kilometres from the central
business district, was originally a mangrove swamp.
In 1978, when the research began, recently arrived settlers were consolidat-
ing the 10 by 30 metre waterlogged plots (solars) they had purchased cheaply
from professional invaders. Households lacked not only dry land, but also
basic physical services such as electricity, running water, plumbing, as well
as adequate social services like education and health facilities. At this time it
was a young population, struggling to make their way in the city, many just
starting families. By 2004, Indio Guayas was a stable urban settlement with
physical and social infrastructure, and, due to the city™s rapid expansion, a
community no longer on the periphery. By this time, children of the original
settlers had reached adulthood and started families of their own, either in the
same community or elsewhere. The study is contextualized within the broader
macroeconomic and political structural context during different phases of
Ecuador and Guayaquil™s history; in brief these can be summarized as the
1975“85 democratization process, the 1985“95 economic structural adjust-
ment policies, and the 1995“2005 globalization and dollarization period.
Three quantitative household surveys were undertaken in 1978, 1992, and
2004 that comprise a panel dataset of the inhabitants of ¬fty-one family plots.
In 1978, a universe survey of 244 households was undertaken over the eleven-
block area; in 1992 a random sample survey of 263 households undertaken in

<<

. 4
( 13)



>>