You are on page 1of 8

Multilevel Modelling of Income and Market Participation

Termpaper submitted to the Dina Research School for the PhD course on Likelihood-
based inference for hierarchical/mied statistical models! "reve! Denmark

b#
$harat P $hatta and Soleiman Mohammadi Limaei
%ugust &''(
)-Introduction
Isolation contributes to rural poverty. Rural life stagnates and local development prospects remain
limited without a minimum of reliable and efficient access to locations of basic social and economic
activities (Schelling and Lebo, 200!. Rural infrastructure is therefore a ma"or priority for reducing
rural poverty (#orld $an%, &&'!.
(arious studies provide evidence that socioeconomic development is closely associated with road
transportation. )overty is more pervasive in areas with no or unreliable road access. *or e+ample, in
,epal, where the percentage of people below poverty line is as high as '2 percent, the incidence of
poverty in areas not connected by roads is -0 percent (Schelling and Lebo, 200!. In $angladesh, the
enrollment of girls in primary schools is three times as high in connected villages compared to
unconnected areas (ibid!. In .ndhra )radesh, India, the female literacy rate is / percent higher in
villages with all weather road access compared to those with unreliable road access (Liu, 2000!. )lenty
of further evidence of the economic impact of rural roads e+ists ()an%a", &&&!. #e therefore attempted
to e+amine the impact of some variables li%e road accessibility , number of o+en, amount of own land
and etc. on household income and mar%et participation in 0igray, the northern highlands of 1thiopia.
0he aim of the pro"ect was to develop a hierarichal model and to analy2e the fi+ed and random effects
about the detrimiation of income and mar%et participation.
&- Data and *stimation Methods
Data and variables
0he data used in this pro"ect were collected from household survey underta%end in 0igray in 200/. /-2
households were selected adopting a stratified random sampling approach based on mar%et access,
population density, rainfall, and irrigation pro"ects from 3 communities located in four different 2ones
of 0igray. 0he data set that we used in the pro"ect contains different socioeconomic varialbes such as
access to information and public goods, household income and asset status, household characteristics,
and mar%et participation. Specifically, the varialbes used in the pro"ect and thier definition are as
follows4
income 5 6ousehold income
road 5 Road accessibility measured by the wal%ing time to get to the
nearest all weather road from the household
hse+ 5 Se+ of household head
hedu 5 1ducation of household head
.ge 5 .ge of household head
7+en 5 ,umber of o+en in a household
7wnland 5 Land owned by each household
8mp 5 8redit mar%et participation by the hh
8wr 5 8onsumer wor%er ratio defined as the household si2e devided
by the number of wor%ers in a hh
9lab 5 9ale labor endowments in hh
*lab 5 *emale labor endowments in hh
0lab 5 0otal labor endowments in hh
0lu 5 0ropical livestoc% units in a hh
6ere household income, denoted by income, and credit mar%et participation by a household, denoted
by cmp, are our dependent variables used in the pro"ect. #e have identified different factors which
determine or are correleated with income and cmp based on theory and empircial studies.
+ierarchial structure! and fied and random effects
)rincipally there are five hierarchies in our data, vi2 from bottom to top household, village, district,
2one, and state. $ut we do not have such clear and distinct hierarchies in our data so we developed our
model in two hierarchies, vi2 household level and village level. 0here are /-2 household level
observations and ' village level observations.
Data anal#sis and soft,are
0he 9Lwi, software was used to analys the data. 0he model constarcted on based of hierarchical data
structure li%e4 village level 3 and houshold level /-2.
Model formulation
0he mi+ed model can be written as
y 5 :; < => < ?
#here y denotes the vector of observed y@s, : is the %nown matri+ of +i"@s , ; is the un%nown fi+edA
effects parameter vector, = is the %nown design matric and > is the vector of un%own randomAeffects
parameter. #e have formulated two models for our pro"ect. 0hey are as follows.
Model I- Income
0he factors determining the rural farm householdsB income may be broadly divided into four
constructs4 mar%et and access to information, household resource endowments, household
characteristics, and village level variables and public goods.
Model II- Market Participation
0he factors influencing or correlated with the rural farm householdsB mar%et participation may be
broadly divided into four constructs4 mar%et and access to information, household income and resource
endowments, household characteristics, and village level variables and public goods. 6ere mar%et
participation by households is a binary variable ta%ing only two values4 either participate or not i.e.
for mar%et participation and 0 for not participation. So the distribution is binomial and the regression is
logistic.#e used the logit model to e+amine the mar%et participation by the households. $ased on
theory and empirical evidence mar%et participation may depend on road accessibility, income, amount
of own land, number of o+en, age, se+ and education of the household head.
.- Results and Discussion
Model I- Income
Model interpertation -
#e have two levels of hierachy in the model4 houshold at the first level and village at the second
level. 0he model we have now set up is a member of the variance copmonent family. $ecause there is
only one set up random effects(intercepts! for each level in the model.
.s we can see from the results of the regaression output (*ig '.! road accessibility, o+en and land
holding significantly contributed more income of the houeholds in the study area. $ut there is no
significant relationship between household income and hse+, and hedu. 0he results also show that there
is high variation of income between housholds than between villages.
/ig0.0)0 9odel of houshold income
Residuals of the data also were calculated in houshold and village lavels (*ig.'.2!.
/ig0.0&0 Residuals of the data on houshold level /ig0.0&0 Residuals of the data on village level
Setting up a random intercept model in the e1uation-
.fter adding random effects on the independent variables we found that there are significant varition
for o+en and owned land in the housholds levels but there are not siginficat varition between the
villages levels. 0here were not also any siginficant varitions on hedu, hse+ and road.
Predictions in the model-
*irst we added a Cuadratic own land in the model, this variable also is strongly siginficant with
coeeficient &3./& and standard error 2/.23.
0he predicted eCuation is li%e4
Income5
0
cons<
6
ownland<
8
(ownland!
2
.
0he predicted eCuation and confidence interval were ploted to see how much uncertainty we had in our
model. . linear prediction was also added to the model to compare the linear and Cuadratic own land
eCuations (*ig.'./!.
/ig0.02- Duadratic, linear and confidence interval for ownland variable.
Model II- Market Participation
*irst we formulated the mar%et participation model ta%ing variables such as se+ of household head,
education of household head, number of o+en, amount of own land , road accessibility as e+planatory
variables but we found that age and se+ of the household head were only significant but other
variables were not siginficant (*ig. '.'!.
/ig0 .0.0 9ar%et participation model
*rom the model we see that the household level variability is more than that of the village level
variability. 0he random effect for age of household head is siginficant but the random effect of se+ of
household head is not significant. #e may conclude from the income and mar%et participation models
described above that the household variability is more than the village level variability.
Residuals of the data also were calculated in village lavels (*ig.'.E!
/ig0.0(0 Residuals of the data on village level at mar%et participation model.
(- 3onclusion
In both income and mar%et participation models, the household variability is more than the village level
variability in determination of household income and mar%et participation. .s we hypothesi2ed, road
accessibility was positively significant to determine the household income. $ut contrary to our
hypothesis Road accessibility was not significant factor in determining the mar%et participation by the
households in the study area. .ge and se+ of hh head were significant factors influencing hhs@ mar%et
participation.

You might also like