Analysis of poverty
and its covariates among smallholder farmers in eastern Hararghe highlands of
Abstract
This paper probes into analysis of the extent and determinants of rural
household poverty in the eastern highlands of
1.
Introduction
A number of studies have sought to examine
the extent of poverty in rural
Bigsten et al (2003), Sharp and Devereux (2004),
Dercon et al (2005) and Little et al. (2006) study the dynamics of
poverty and consumption levels and emphasize the role of shocks in influencing
fluctuations in consumption levels over relatively short periods of time.
Dercon et al (2005), based on study on consumption levels in 15 Ethiopian villages
in the period of 1999-2004, found that virtually all households experienced
adverse effects of shocks, among which they enumerate drought and illness as
most important. The former reduced consumption levels in the sample by 20% and
the latter by 9%. Policy shocks, such as risk of land distribution or arbitrary
taxation were found to be less significant – a change compared to earlier
studies conducted (Dercon, 2001). Investigating poverty dynamics in South Wollo
between 1997 and 2003 (including the 1999/2000 drought), Little et al (2006)
found that the incidence of poverty in their survey area changed little, but
the very poorest improved their welfare a lot, though not sufficiently to
escape poverty.
Research on factors that affect incidence and dynamics of rural poverty in
This paper studies household and
community level covariates that affect the probability of a rural household to
be poor (on various levels of poverty or deprivation) at a particular time. The
study is based on survey of 216 households in three districts in the Eastern
Hararghe highlands of
2.
A brief description of the study
region, sample design and data
Eastern Haraghe highlands are among the densely populated regions of
The most important perennial crop in
the zone in terms of area coverage is khat.
Khat is a shrub which is grown for
its narcotic substance found on the tip of the young leaves and chewed when
fresh. Khat requires the least input
in terms of labour and fairly resistance to pests and disease and withstand
moisture stress. It is common to observe sorghum and maize in intercropping and
also intercropped with khat.
Some of the areas in the study
region have access to regional markets (Harar, Dire Dawa) and road
infrastructure which encourage a market-oriented production system, for example
vegetables and khat. The latter is
grown and marketed domestically and internationally, often through contraband
trade into
The analysis of poverty in this
paper is based on a household survey conducted in three districts in eastern
Since the sample
represents a group of households in specific districts in one particular region
of
Table 1 provides
demographic characteristics of sample households in the three districts.
Household size and age composition of the household determines labour
availability and levels of dependency of younger members. The average age of
the household head in the sample varies from 36.48 years in Kombolcha district
to 42.81 years in Babile. A comparison across districts using a statistical
F-test at the 95% standard confidence interval showed no significant
differences in household characteristics except age of household head. 12.5% of
the sample households are female headed.
Table 1 about here
Table 2
illustrates household asset endowments across the three districts. These
figures show that the average land holding size is 1.37, 0.77 and 0.61 hectare
per household in Babile, Kersa and Kombolcha districts respectively. Babile
district has the largest household size and the largest land holding per
household. But Babile is located in the semi-arid and is drought prone with
marginal agro-ecology for sedentary farming. Households in this district are
among the poorest in terms of income and consumption. Kersa and Kombolcha
districts are considered better agricultural areas with high population
densities. In these districts, soil degradation due to continuous expansion of
the agricultural frontier towards less favourable locations is one of the major
challenges. Nevertheless the latter two districts display higher levels of
income and consumption, have better access to infrastructure, markets and education.
In these areas, cultivation of khat
as high value cash crop has become more and more pervasive.
Table 3
summarizes livestock ownership by sample households. In the highlands of
Tables 2 and 3 about here
3.
Analysis of household poverty
The analysis in this
section is based on the expenditure dataset of the sample households. Household
expenditure is considered as an adequate measure of household welfare in
developing countries as it is better able to capture household’s consumption
capabilities (Grootaert, 1986). There are two main reasons to study consumption
expenditure as compared to net earnings from various livelihood activities.
First, some components of household consumption are usually measured more accurately
than income, and second, consumption is less susceptible to income volatility,
especially in the context of rural households in developing countries which
strongly depend on agricultural income.
A household is
considered as poor when household expenditure is insufficient to meet the food
and other basic needs of all household members. To make the assessment, a
basket of goods and services corresponding with local consumption patterns and
satisfying a pre-set level of basic needs for one person is constructed and
valued at local consumer prices to compute its minimum cost. The value of this
basket is called the “poverty line”, and is most appropriate if expressed in
per-adult equivalent terms. If the per-adult equivalent expenditure of household
members is found to be below the poverty line, the household and its members
are considered poor, non-poor otherwise (Lipton and Ravallion, 1995). Even
though the data requirements of this method are very steep, and very comprehensive
questionnaires are needed to collect it, it remains to be a widely accepted
measure of poverty—as far as its economic dimension is concerned. In this study
too, household expenditure on basic needs - including those on food, clothing,
housing, education and medical care-is used as a measure of household welfare
(Glewwe, 1991; Alderman, 1986).
In this study, we follow the
Foster, Greer and Thorbecke (Foster et al., 1984) class of poverty index. Let
us begin with some notations. Define a vector of a suitable measure of living
standards Y (household calorie intake
per capita, or expenditure) in increasing order, Y1, Y2, Y3, ..., Yn,
where n represents the number of
households under consideration. The General Foster, Greer and Thorbecke (FGT)
poverty index (Pa) can be expressed as:
(1)
where:
x represents income; z is the poverty line; q is the number of the poor; gi
is shortfall in chosen indicator of standard of living, say expenditure
per capita shortfall of the ith household. That is, let xi denote
the per capita expenditure of household i, then gi =z-xi
if xi < z ; gi =
0 if xi ³ z. a represents poverty aversion parameter (measure
with larger a are more sensitive to the poorest
of the poor. For a = 0, Pa will be equal to the poverty
headcount ratio; for a = 1, Pa will be equal
to the normalized poverty gap and for a = 2, Pa will be equal to the squared
normalized poverty gap ratio.
The headcount ratio (a = 0) does not tell us whether the
poor are only slightly below the poverty line or whether their consumption
falls substantially short of the poverty line. Moreover, the head count measure
also does not reveal whether all the poor are about equally poor or whether
some are very poor and others just below the poverty line. Two regions may have
the same head count ratio but the poor in one region may be much poorer than
the poor in the other one. The poverty gap (a = 1) partly overcomes these
problems by incorporating the depth of poverty. It is a poverty measure that
takes into account how far the poor, on average, are below the poverty line.
That is, each poor person is weighted by his/her proportionate shortfall below
the poverty line indicating how poor he/she is.
When setting a equals to 2, we obtain the squared normalized
poverty gap ratio, often called the severity of poverty or FGT(2). This poverty
index gives greater emphasis to the poorest of the poor by weighting each poor
person by the square of his/her proportionate shortfall below the poverty line.
P2 is more sensitive to redistribution among the poor in that a dollar
gained by the very poor would have more effect on poverty as that gained by the
moderately poor. P2 is more comprehensive because it increases when
the number of poor people increases, or the poor get poorer, or the poorest get
poorer compared with other poor people (Greer and Thorbecke, 1986).
A feature common
to all poverty literature is a significant degree of arbitrariness in the value
assigned to the poverty standard. Even with those approaches based on
subsistence needs, the absence of one level of food intake required for
subsistence, rather a broad range of combinations, makes constructing a
suitable poverty index more complex. Ideally, the poverty line should be based
on a basket of goods and services including food and nutrition, as well as clothing,
housing, health care and education that can be considered basic needs (Baffoe,
1992).
In the absence
of an invariably acceptable national poverty line for
Moreover, the
Gini coefficient will be used to measure the degree of inequality in
expenditure and income. Mathematically, the Gini coefficient can be directly
defined by:
(2)
where there are n households indexed by i,
their respective per capita household income is given by yi , mean per capita household income is
denoted by m and ri is the rank of household i in the y-distribution, counting from
the bottom so that a household with the greatest per capita household income
has the rank n. Using equation (2), the Gini index can be straightforwardly and
rapidly calculated from the household income, expenditure or any welfare
indicating data after sorting the observations. According
to Gini coefficient inequality can vary from 0 to 1. When the Gini coefficient
is equal to zero, income is fully equally distributed. When Gini coefficient
approaches to one, income is extremely unequally distributed. The distribution of total household
expenditure, expenditure per capita and expenditure per adult equivalent
illustrates disparities across districts surveyed and also between the poor and
the non poor.
Table
4 about here
Comparing the three
districts in our study region (Tables 4, 5 and 6), we observe that households
with the lowest expenditure live in Babile district and spend approximately 88
percent compared to a typical household in the study area both in terms of total
household expenditure and expenditure per adult equivalent. As it is apparent
from Table 5, poor households in Babile district managed to spend only 45
percent of their counter part non poor households in the same district, where
as the corresponding proportion for poor households in Kersa and Kombolcha
district are 59 percent and 57 percent respectively. Furthermore, we find that
poor households on average spend approximately 68 percent and 51 percent in
terms of total household expenditure and per capita expenditure as compared to
the non-poor households. This diversity in expenditure level reflects the
heterogeneity between the poor and the non-poor and also between districts,
which are considered to be highly vulnerable for resource degradation and others
with relatively better potential for agriculture.
Table 5 and 6 about here
Based on the poverty
line estimated earlier, the analysis undertaken for the whole sample households
yields a poverty head count ratio of 0.356, that is, 35.6% of the total
population spends less than what they would need to meet minimum living
standard requirements. By decomposing across districts, we observe a
differentiated picture of the distribution of poverty. Tables 6 depicts that a
high proportion (52.3%) of the population in the Babile district lived under
poverty in 2003/04 followed by Kombolcha district with head count ratio of
0.301. The results with respect to the depth of poverty and severity of poverty
also show that both the depth and severity of poverty seem to be highest in
districts with highest incidence of poverty. One can observe that not only is
the incidence of poverty in Babile district the highest, almost three times
that of Kersa and two times that of Kombolcha, but also poverty in Babile is
found to be deeper and more severe
The Gini
coefficients for the sample households are found to be 0.23 and 0.29 for
expenditure per adult equivalent and income per adult equivalent. These
coefficients are relatively low and suggest that overall inequalities in expenditure
and income per adult equivalent among the sample households are not severe.
Inequality in expenditure per adult equivalent also varies among the districts.
The results indicate that expenditure is more unequal in Babile district as
compared to the other two districts under consideration.
4.
Covariates of rural poverty
While economic theory provides a well-developed framework for studying
earnings and income dynamics, no similar and uniform theory exists that could
guide us in the more complicated case of poverty dynamics (Bigsten et al.,
2003; Barrett and Swallow 2006; Glauben et al., 2006; Dercon, 2006). In
principle a whole variety of factors could be considered as important
determinants of lifetime poverty, among those are: age, human capital (formal
or informal education), sex of household head, household size, and resource
endowment. Age, in most cases, is hypothesized to be positively correlated with
well-being. However, a negative relationship can also be explained by the
assumption that as farmers grow older, there is an increase in risk aversion
and a declining interest in long-term investment in the farm. Younger farmers
tend to be less risk-averse and are more willing to try new technologies, which
may lead to better income.
Education is perhaps
the strongest correlate of poverty, insofar as it determines the command of
individuals over income earning opportunities through access to employment.
Education was typically found to have a high explanatory power to observed
patterns of poverty. The correlation between education and welfare has
important implications for policy, particularly in terms of the distributional
impact. Formal education increases the farmer’s ability to understand and
respond to information. Human capital increases the ability to think analytically,
make practical adoption decisions, and use a technology appropriately to
increase family income (Weir and Knight, 2004). Studies show that the educational attainment of the head of the
household is an important factor that is associated with poverty (Alemayehu et
al., 2001). Lack of education is a factor that accounts for a higher
probability of being poor.
Moreover,
dependency ratio and number of persons per household tend to be correlated
negatively with the probability of being poor. The effects are expected to be
stronger for females than males. That is, we expect that the greater the
household size, the higher the probability that any particular household is
classified as poor. However, households containing members able to participate
in on-farm and non-farm activities can enable farmers to adopt labour-intensive
technologies (Feder et al., 1985). If technologies are capital-intensive,
household members may work non-farm to generate more income. Property related
characteristics of households including farm size and livestock holding can
potentially determine poverty of households. In general, populations with
higher non-farm incomes exhibit a willingness to accept more risk and adopt
complex technologies. Farmers with larger farms can invest more in information
acquisition and accumulate knowledge that leads to better return.
A particular
interest in our study is the role of active membership in various types of
local level organizations and networks – often referred to as “social capital”
(Putnam 1993; Woolcock and Narayam 2000; Donnelly-Roark et al.,. 2001;
Grootaert and Van Basteler 2002) - as a covariate to household poverty. We look
at different types of local-level networks and organizations, namely (a)
governance and administrative networks (GALLI), e.g. involvement in peasant
organizations, (b) social and religious (SRLLI) and (c) natural resource and
productive networks and organizations (NPLLI), e.g. involvement in networks of
labour exchange. Table 7 presents code, definitions and descriptive statistics
of variables considered in the empirical analysis.
Table 7 about here
4.1.
The empirical model
Qualitative response models are often used when a dependent variable
takes one of a number of discrete values. Such models estimate the
probabilities of being poor using Maximum Likelihood Estimation (MLE) while
accounting for the discrete nature of the dependent variable (Greene, 2002).
Binary response models (e.g. probit, logit) are used where poverty is considered
as a “yes” or “no” decision. However, in this study we do not only consider
whether a household is poor or not, but also the intensity or depth of poverty.
Therefore, the model needs to consider more than two possible responses.
Similar approach has been followed by Alemayehu et al. (2001) which focus attention on the hard-core poor, moderately poor
and non-poor categories to employ an ordered logit model. This approach is
justifiable, because it explicitly makes the ordering of the population
sub-samples, using the poverty lines as cut-off points in a cumulative distribution
of expenditure. Whenever poverty categories have a natural order, the ordered
probit is the appropriate model to be employed in the estimation of relevant
probabilities (Greene, 2002). Ordered response models recognize the indexed nature of various response
variables. Underlying the indexing in such models is a latent but continuous
descriptor of the response. In the ordered probit model, the random error
associated with this continuous descriptor is assumed to follow a normal
distribution.
The ordered
probit model differs from a univariate probit one in that the dependent
variable is no longer a dummy variable, but an ordered variable taking values
0, 1, 2, 3 according to the level of poverty the household is encountered with.
As in a univariate probit model, the model is built around a latent regression
variable. An ordered probit model allows for multiple ordered values for the
dependent variable and analyzes the effect of each independent variable on the
dependent variable. It measures the probability that this dependent variable (Yi , for the i-th household) falls in
one of the discrete categories conditioned on levels of the independent
variables(Xj). Suppose the level of poverty of the sample household
i (Yi*) is the unobserved
variable (latent variable) and Yi*
is expressed in the following equation:
(3)
where xji
are the above mentioned explanatory variables; ui are the residuals
or error term and the β and µi are parameters to be estimated
(Greene, 2002). We assume that ui is normally distributed across
observations. As mentioned previously, Yi*
is unobserved and we can only observe whether the household under consideration
falls in category “0,” “1,” “2,” or “3”.
So, what was observed is the following actual placement in the discrete
category:
Yi = 0 if Yi*
< µ1 (extremely poor)
Yi = 1 if µ1
≤ Yi* < µ2 (moderately poor)
Yi = 2 if µ2
≤ Yi* < µ3 (slightly non poor)
Yi = 3 if µ3
≤ Yi* (non
poor)
In this model, Y
(the dependent variable) represents the intensity of poverty experienced by a
household. Here intensity of poverty is defined according to the following four
categories:
0 = extremely
poor; PCAE[2]
expenditure less than Br. 1102
1 = moderately
poor; PCAE expenditure lies between
2 = slightly non
poor; PCAE expenditure between
3 = non-poor;
PCAE expenditure more than Br. 1835.
Coefficients of the ordered probit model (β) give an indication of
positive or negative impact of an independent variable on the probability of
being poor, but do not relay information concerning the magnitude of the
effect. Using a transformation function, the model creates a linear index of
the probabilities with a cumulative standard normal distribution. Given the
classification, we can derive the probabilities of being poor of different
degrees as follows:
Pr(Yi = 0)
= Pr(Yi* < µ1) =
Φ(µ1 – β’Xi)
Pr(Yi = 1)
= Pr(µ1 ≤ Yi* < µ2) = Φ(µ2 – β’Xi ) - Φ(µ1 – β’Xi)
Pr(Yi = 2)
= Pr(µ2 ≤ Yi* < µ3)
= Φ(µ3 – β’Xi) – Φ(µ2 – β’Xi)
Pr(Yi = 3)
= Pr(µ3 ≤ Yi*)
= 1 - Φ(µ3 – β’Xi)
where µi
represent the threshold or cut-off parameters for placement of Yi* in the discrete poverty
categories, and Φ( ) is the standard normal cumulative distribution function
such that the sum total of above probabilities is equal to one. We maximize the
log-likelihood function to obtain the estimates of µ’s and β’s employing
LIMDEP statistical software.
Marginal effects
are calculated using the linear probability index. They tell us the effect on
the probability of being poor in a particular category for changes in the
independent variables (∂Pr(Y=0, 1, 2, and 3)/ ∂Xi). The marginal effect is
the percentage change on the probability associated with a unit change in the explanatory
variable. The marginal effect for each variable is calculated at the mean
values of the independent variables. In this context, it is possible to assess
the probability of being poor for given factors, and comparisons can then be
made across characteristics.
Since there is a
debate in the literature as to whether it is better to estimate an ordinary
least square model using continuous expenditure data or use the categorical
poverty level, we will complement the ordered probit model with OLS
estimates.
4.2.
Results and discussion
The use of an ordered probit model enabled us to look at how particular
variables affect the extent of household poverty. The results of the ordered
probit estimation presented in Table 8 depict that the signs of most of the
estimated parameters conform to our expectations with the exception of TLHPAE
and TLU. But both were statistically insignificant (P>0.10). The likelihood ratio
test for the goodness of fit shows a good fit for the model (P < 0.001).
In general, nine
of the fifteen variables were found to be statistically significant in the
ordered probit model at less than 10% probability level. Among the nine
statistically significant explanatory variables, we found age of household
head, non farm income, proportion of irrigated land owned, active participation
in productive and social local level institutions and residence in Kersa and
Kombolcha districts to be positively related to household well-being. Whereas
size of household in adult equivalent and active membership of natural resource
related local level institutions are covariates that are negatively correlated
with the probability of being non-poor.
Given that the
dependent variable of our regression, ORDPOV, is an ordered variable, we
calculate the marginal effects of a unit change in a number of explanatory
variables for the four categories of poverty which, to some extent, would
reflect the effect of a unit change in any explanatory variable on the probability
of a household of being extremely poor (ODRPOV = 0), moderately poor (ORDPOV =
1), slightly non poor (ORDPOV =2), and non poor (ORDPOV = 3). Table 9 shows the
estimates of marginal effects of the variables, which allow further assessment
of the estimate with respect to each poverty category. These marginal effect
figures further strengthen the inferences obtained from the parameter estimates
in the ordered probit model. In particular, we focus on the marginal effects
which are statistically significant in determining household poverty status,
namely age of the household head, size of household in adult equivalent, non
farm income, active membership in local level networks and organizations and
the location.
Table 8 about here
Age (to a limit)
is expected to be associated with skills enhancement (experience), accumulation
of resources, extensive social capital and others that ought to contribute
positively to well-being (Bashaasha et al., 2006). Our results seem also to
confirm this statement. Age of household head is found to be positive and
statistically significant (p < 0.10), implying that among the sample households
older households have greater likelihood of being non poor. More specifically,
an increase in age of household head by one year would increase the probability
of being slightly non poor and non poor by 0.11 and 1.64 percent, respectively,
where as it lowers the likelihood that a household will fall under category
extremely poor and moderately poor by 0.66 and 1.09 percent respectively.
Family size reflects the number of units among which household resources need
to be allocated according to the weights of each unit. Family size may have an
ambiguous role in poverty status of rural households depending on the relative
strength of size economies in consumption as against the diminishing return to
scale. In our sample, increase in household size by one adult equivalent would
increase the probability of being extremely poor and moderately poor by 3.13
and 5.16 percent, respectively, where as it lowers the likelihood that a
household will fall under category slightly non-poor and non-poor by 0.49 and
7.79 percent respectively.
Access to a
non-farm source of income is also an important determinant of wellbeing in
eastern
Results of the
ordered probit model indicate that better off households are more likely to
participate in social and religious (SRLLI) and governance and administration
(GALLI) networks and organizations, where as the poorer are active with natural
resource and production related networks and organizations (NPLLI). This result
finds its explanation from the fact that natural resource related local level
networks are largely supported by NGOs and also coordinated by the district
bureau of agriculture so that rural households can participate in conservation
practices such as building and maintaining terraces, planting trees and
construction of feeder roads in return for food items through food for work
programs. In this sense, participating in the latter offers immediate benefits
in the form of food for poor households, but not necessarily social assets upon
which further networks of social and economic benefits for the future could be
built. This finding indicates that poor households are significantly
underrepresented in governance networks as well as social networks. These
networks are dominated by non-poor households.
Two district
dummies for the three districts accounted for location-specific, district-level
variations in the provision of public services, market opportunities and
vulnerability to ecological uncertainties across the study districts. The
probability of being non-poor was 21.40 percent for a rural household living in
Kersa district, but only 17.55 per cent in Kombolcha district.
In order to
scrutinize whether the ordered probit model has suffered from loss of
information in the process of categorizing the dependent variable, we estimated
an ordinary least square model (OLS) with continuous expenditure data as
dependent variable whereas the explanatory variables remaining the same. The
result indicates that (i) all the variables have the same sign except the dummy
for Kombolcha district, (ii) only six explanatory variables turned out to be
statistically significant with OLS estimate as compared to nine variables for
ordered probit model and (iii) age dependency ratio turned out to be
significant with OLS but not with ordered probit model.
5.
Conclusions
This paper has studied extent of and the determinants of poverty in
three districts in rural areas of eastern
Our study points, among others, to three main reasons that explain the extent and variation in poverty levels across households studied: (1) poverty is location-specific as the stark variations between Babile and the other two districts has shown. This indicates how endowments with market access and relatively better agro-ecological conditions are essential factors in increasing household welfare, something where outside intervention can only partly help improve the situation. (2) Access to irrigated land (not land per se) and non-farm income are strongly correlated with lower probabilities of being poor. (3) Involvement in networks is a strong predictor of the probability of being poor – and we identified a clear differentiation in the types of networks that matter. Whereas poor households tend to participate in externally driven natural resource management networks, often induced through food for work incentives, the networks that really impact upon poverty levels are governance and social networks. It appears that active membership in the latter two is strongly correlated with a lower probability of being poor. This indicates that poor households face some kind of exclusion from those networks, possibly because others intentionally exclude them or because they cannot afford to par