African Census Analysis Project (ACAP) UNIVERSITY OF PENNSYLVANIA

Population Studies Center 3718 Locust Walk Philadelphia, Pennsylvania 19104-6298 (USA)

Tele: 215-573-5219 or 215-573-5169 or 215-573-5165 Fax: 215-898-2124 http://www.acap.upenn.edu Email: tukufu@pop.upenn.edu

Analysis of Regional Differentials in Under-five Mortality in Kenya Using a Count-Data Regression Model

Gideon Rutaremwa

ACAP Working Paper No 14, March 2000 This research was done as part of the African Census Analysis Project (ACAP), and was supported by grants from the Rockefeller Foundation (RF 97013 #21; RF 98014 #22), from Andrew W. Mellon Foundation, and from the Fogarty International Center and the National Institute of Child Health and Human Development (TW00655-04). We would like to thank Timothy Cheney for computer programming assistance.

Recommended citation:

Gideon Rutaremwa. 2000. Analysis of Regional Differentials in Under-five Mortality in Kenya Using a Count-Data Regression Model. ACAP Working Paper 14, March 2000, The African Census Analysis Project (ACAP), Population Studies Center, University of Pennsylvania, Philadelphia,, Pennsylvania.

Abstract

ines regional

he number of

tive binomial

sults showed

higher incidence rate ratios of under-five deaths for Nyanza and Western Provinces,

while lower rates were estimated for Central and Nairobi Provinces. These regional

ls were added

ortality exist

tality and the

anatory variables namely: age of mother, her education, marital status, type of toilet

facility, source of water and place of residence, is basically similar in all the provinces of

Kenya.

Using data from the Kenya 1989 population census the study exam

differences in under-five mortality using a count-data regression model. T

children dead for each woman is used as the dependent variable in a nega

regression equation to explore regional under-five mortality differences. Re

differences in under-five mortality were reduced in magnitude when contro

in the regression models. Nevertheless, regional differences in childhood m

in Kenya and are substantial, and the relationship between under-five mor

expl

ii

Introduction

Understanding the geographic distribution of mortality is crucial to policy

Africa tends

this paper I

a, taking into

consideration various environmental and socioeconomic factors. Among the reasons for

regional differences in child mortality include geographic conditions such as climate and

altitude, which may affect land productivity, especially in areas where the level of

also increase

and Knowels,

evant variable

for mortality differential among children under the age of three. In addition differential

access to political power has been found to explain demographic behavior patterns in

Kenya (Weinreb, 2000).

rtality in East

s the method

ich a mortality indicator is generated.

Th ble in a linear

regression model. However, these procedures rely on normal assumptions, which may

not be accurate in under-five mortality covariate analysis.

In the current study, data on reports of women concerning all their live-born

children who have subsequently died are used to estimate a negative binomial model.

This model derives from a Poisson distribution, which has also been described as the

intervention (United Nations, 1991). Mortality in most parts of sub-Saharan

to cluster by area, often identified as a high or low-mortality region. In

examine the spatial distribution of under-five mortality using census dat

technology in agricultural production is still low. Climatic conditions can

the incidence of infectious and parasitic diseases. Studies on Kenya (Anker

1977; Ewbank et al., 1986), have identified endemic malaria as the most rel

There is a large empirical literature on covariates of under-five mo

Africa. The approach used in several studies (United Nations, 1991) i

proposed by Trussell and Preston (1982), in wh

e mortality indicator, once computed, is then used as a dependent varia

1

benchmark model for count data analysis (Cameron and Trevedi, 1998, Allison, 1998,

Long 1997). Although Poisson models have been widely used in applied econometric

rial science,

any of the

rogeneity and

all have maintained the relevance of these models in the context of count regressions.

The standard generalized form of the Poisson model is the negative binomial model,

tudy because

e conditional

dent variables are used as

predictors including maternal education, marital status, region of residence, rural/urban

residence, occupational status, source of water, and type of toilet facility.

ge number of

which makes

me regional

lity. Brass and Jolly (1993) suggest the need

for greater geographic disaggregation in examining levels and trends in childhood

mortality. The relevance of such regional-level analyses of patterns, trends, and

differentials of mortality cannot be overemphasized.

The current study focuses on some of the issues that are relevant to understanding

regional differences in under-five mortality in Kenya. In so doing, the study attempts to

works, early applications of the Poisson model took place in actua

biostatistics, and demography (Cameron and Trevedi, 1998). M

developments in analysis of count data have addressed such issues as hete

overdispersion as well as true versus apparent contagion in Poisson models, but have

derived by Greenwood and Yule (1920).

The negative binomial regression model is useful in the current s

of the role of covariates that are thought to affect the parameters of th

distribution of deaths among children. A number of indepen

The analysis uses census data as opposed to surveys because a lar

cases are required for adequate analysis of mortality at the regional level,

surveys unsuitable. Census data also offer an opportunity to examine so

covariates and patterns of under-five morta

2

answer the following questions: first, are there substantial regional variations in under-five

mortality? Second, what factors are most related to under-five mortality in the different

finally, do these differences disappear once individual and

hou

e 1989 Kenya

cluded in the

censuses. For example, census data do not contain information on causes of death,

immunization, access to and use of health facilities, and nutritional status in the various

regions of the country, yet these factors are known influence levels of under-five mortality

ries.

The study considers 135,459 Kenyan women of reproductive ages (15-49) and

who had borne at least one child by the time of the census. These data contain certain

ity in child mortality

incl r province of

, and marital

r.

It is important to note that the data used for this study have certain errors; notably,

about 15 percent of the women in the Kenyan sample did not state their parity. The

majority (about 60 percent) of these women were aged 15-19 and about 71 percent of

them were single. Given the age and marital status distribution of women in the Kenyan

sample, it seems reasonable to assume that there is little impact of these missing data on

regions of Kenya? And

sehold factors are controlled?

In an attempt to provide answers to these questions, this study uses th

census data. It is important to note that some of the relevant questions are not in

in developing count

Data and Methodology

variables that are important in analyzing geographical divers

uding information on children ever born and those surviving, region o

residence, district of residence, rural or urban residence, educational status

status. Household-level data include type of toilet facility and source of wate

3

the overall reporting of children ever born and surviving and also on child mortality

estimates derived from these data.

Cha

e presented in

Table 1. Overall, less than 10 percent of the respondents were in the age-group 15-19, this

proportion increases to slightly above 20 percent in the age groups 20-24 and 25-29, and

45-49. Age is

erspective, in

untries births to older women are usually of high order and are prone to higher

mor oung mothers

Results in Table 1 also suggest that about 20 percent of the Kenyan population are

rning level of

ment, Nairobi as well as Central Province had the highest proportion of

wom y about 1

North Eastern

The marital status variable shows that overall over 80 percent of the study population

was currently married. The expectation is that women who are currently married have more

support from their partner and are in a better position to look after their children well,

resulting in higher child survival compared to those not in a marital union. Results in Table 1

show that Provinces that were mainly rural had higher proportions of married women.

racteristics of the Respondents

Data on some of the characteristics of the respondents in this study ar

subsequently declines gradually to less than 10 percent in the age groups

important because it is closely related to child survival. From a theoretical p

developing co

tality risks. This is also true of first-order births that occur to very y

(Rutstein, 1992).

urban. There are regional variations in proportion of urban population. Conce

educational attain

en with a secondary and higher education while the North Eastern had onl

percent within this category. The worst educational indicators are observed in

and Coastal provinces.

4

Results in Table 1 also show the sanitation indicators, source of water and type of

toilet facility. Better sanitation was most lacking in North Eastern, South Rift and Western

Province had

y

or standard of

environmental

sanitation in these households. In addition, they show variation in the level of regional

development, since we expect more developed regions to have higher levels of infrastructure

t facility are

ldren in these

services that could benefit child health are

imp ilities for safe

in this paper.

with human,

that either aid

. For example,

Ewbank and others (1986), explaining regional differences in infant mortality in Kenya, point

to the fact that the high level of mortality observed for the Coastal and Nyanza Provinces

could have been partly due to malaria infestation in these regions. However, other factors,

such as political influence on access to health and other development infrastructure, may be

related to different levels of mortality (Weinreb, 2000).

Provinces of Kenya: approximately 80 percent of women from North Eastern

no toilet facility. Source of water and type of toilet facility, each with three dumm

categories, are included in the analysis because not only are they proxies f

living of the households of the respondents but they are also indicators of

use. Respondents from households where pipe-borne water and flush toile

available enjoy a higher standard of living, and child survival among chi

households is higher. Among the basic

rovements in the quality of the drinking water and the provision of fac

disposal of human excreta (Esrey and Habicht, 1986).

Finally, region of residence is examined as an explanatory variable

Each of the nine provinces of Kenya is expected to be diversely endowed

environmental and other resources that can promote child survival. The factors

or otherwise reduce child survival may be different in the respective provinces

5

Table 1: Selected characteristics of the sample women by region of residence and other selec eted background variable, K nya 1989

We otal

3.4

5.786.5

0.19.9

7.80.6

3.511.1

19.7

2.6

76.0 9.8

13.

6.84.6

86.13.2

22.

12.9.8

32.3

26.0

10.7

684.6

.4

8514.7

19.9

1411.4

.3

44.6

14.5

15.5

5.188.2

7.0

20.7

13.311.1

15.0

15.4

7 5 8

7 2 0

7 3

9 6 5 4 7 2 7

3 2 5

7 6

41

1

8

81

22111

7.3.1.8.

2

5

Region of residence Selected variable/ category Nairobi Central Coastal East North East Nyanza N. Rift S. Rift st T

Education None Primary Secondary +

Marital status Never married Previously married Currently married

Place of residence Rural Urban

Age group 15-19 20-24 25-29 30-34 35-39 40-44 45-49

Toilet facility None Other Flush

Source of water Other safe Other-unsafe Tap/ pipe-borne

12.5 40.9 46.6

21.3 5.7

73.0

- 100.0

5.5

23.7 28.4 17.7 12.2 7.7 4.8

1.8

35.4 62.8

1.0 2.1

96.9

21.4 54.7 23.9

17.5 5.8

76.7

91.5 8.5

4.1

21.8 23.2 16.0 13.8 11.8 9.8

0.6

93.2 6.2

13.0 53.2 33.8

62.5 27.2 10.3

8.8 8.5

82.8

68.4 31.6

5.9

18.4 22.5 19.2 14.5 11.2

8.2

39.5 51.1

9.4

14.7 32.5 52.8

40.9 45.4 13.6

12.0 6.3

81.7

94.2 5.8

4.1

19.4 22.2 17.7 14.5 12.6 9.5

24.3 73.6 2.1

22.3 58.0 19.7

95.9 2.9 1.2

1.7 8.2

90.1

76.8 23.2

3.5

18.2 22.9 20.8 13.0 14.6 7.0

79.0 19.7 1.3

42.5 45.2 12.3

40.6 45.9 1

7.8

9

220.5 17.2 1

9.4

77.8

14.3

41.7 44.3

9

13.4 0

8

7.7

1 22.4 16.7

9

8.5

61.7 6.0

11.8 62.2

56.5 32.8

9.0

.3

5.4

22.5 17.8

8.7

51.5 3.9

33.7 51.8

43.4 41.1

6.7

93.0

6.9

21.0 18.2

8.9

83.2 1.8

49.3 35.3

40.2.6.

1.6.2.

1.8.

5.0.2.

3.68.8.

18.2.

28.7 TOTAL (N) 9476 20454 12879 24838 2456 24096 20499 12626 8135 135459

6

Model Specification and Estimation

equently died

th the various

covariates. Let us assume that a discrete random variable Y (number of children dead)

is Poisson-distributed with intensity or rate parameter µ, µ>0, and t is the exposure,

defined as the length of time during which the events occur. Y is defined by the

following density distribution function:

Data on reports by women of all their live births who have subs

are used to estimate under-five mortality incident rate ratios associated wi

... 1,2,)(e yµt

===− tµ

0,y ,y!

y]Pr[Y

Equality of the mean with the variance is kn

Where E[Y], the expected value of Y, equals the variance - V[Y] = µt.

own as the equidispersion property

of eal-life data.

d Rodriguez

In this study, the dependent variable Y is the count of the number of children

born alive who have subsequently died for mother i, i=1, 2, 3,…, n, where n denotes

the sample size. The count-datum y s distribution depends on a set of exogenous

t ui represent

t:

the Poisson model. This property is frequently violated in r

Overdispersion means that the variance exceeds the mean (Trussell an

1990; Long 1997; Allison 1998a; Cameron and Trevedi 1998).

i

i

variables, some of which are observed (the xi) and some unobserved. Le

unobserved variables and measurement errors on the data and le

{ } λ)u,β,λ(xu,x|YE iiiiiii ==

Where E stands for the expectation operator, β is the k-dimensional parameter vector to

be estimated and u is the unobserved variables and measurement errors in the data.

e:

Implicitly the latter equation assumes that all individuals with the same

characteristics Xi have a Poisson distribution with the same mean. Suppose the ui

(source disturbance not included in Xi) was observed; we could work with the

i

The general form of the log-linear regression model specification would b

uβXuβXlog λk

1kikikiii ∑

=

+=+=

7

unconditional distribution of β given ui, which is Poisson with the m

However the ui is not observed and

ean of λi.

may not even be observable and we are forced to

con

ability of her

of mortality,

hence children ever born and duration of exposure to the risk of dying. This then

ntered in the study.

Fur

ying.

model as an

offset variable. By including ln[Children ever born]=Ψ as an offset in the equation, it

is differentiated from other coefficients in the regression model by being carried

through as a constant and forced to have a coefficient of 1.0. The final model that is

sider the unconditional distribution of βi.

To proceed, we assume that for each individual mother, the prob

children dying depends on the number of children exposed to the risk

allows us to control both the duration of child exposure to the risk of mortality and the

number of children exposed to the risk for a given woman e

thermore, we include age group of mother in the regression models as one of the

covariates in order to account for the age-varying exposure to the risk of d

The logarithm of children ever born is introduced in the regression

estimated is therefore the following:

σε)Xβ...,Xββ(βiD jij2i21i1 ++++ X

0e iψ=

Wh i i hildren born,

β is the vector parameters affecting under-five mortality levels while Xs are the

inear models

case of error or

stoc ectation of the

dependent variable and the linear predictor is a logarithmic function and the linear

of maximum

tatistics.

The model suggests that both sets of parameters are dependent on the covariates.

Furthermore, the number of children born will be equal to the observed deaths if the

coefficients of the independent variables, denoted by β, are equal to zero. Since Ψ is a

constant, any variation in the coefficients of the independent variables will show up

affecting the dependent variable and not the number of children born. The procedure

ere D is number of children dead, ψ is the logarithm of the number of c

covariates of interest.

This final model falls within the framework of generalized l

described by Nelder and Wedderburn (1972), representing a special

hastic structure, which is Poisson-distributed. The link between the exp

predictor contains a known part or offset. This allows for the estimation

likelihood, standard errors, and likelihood ratio goodness-of-fit chi-squares s

8

therefore allows us to obtain maximum likelihood regression coefficient

easily interpreted in terms of differentials in the dependent variable. Using

binomial regression procedure, several regression equations are estimated

relationship between under-five mortality changes when control var

s that can be

the negative

to show the

iables earlier

men

pressed on a

l scale, thus

interpretation of the parameters (β) obtained from the negative binomial regression

models are in terms of incident rate ratios. The incident rate ratios are obtained by

exponentiation of the regression coefficients, that is, exp[β]. For ease of interpretation,

xpression 100*(exp[β]-1) tells us the percentage change in the incidence or risk of

und

The findings from the regression models are presented in Tables 2 and 3. The

results suggest wide and significant regional variations in child mortality in Kenya.

ts change in

mortality was

nza Province

ovince for all

the models fitted to the data and was least for Central Province. According to Table 2,

results in Model I further suggest that the risk of child mortality was 17 percent lower in

the Central region compared to Nairobi. These regional differences in the risk of child

mortality between the Central and Nairobi regions widen when more controls are

introduced in the model. In Table 2 (Model I), results show that the risk of child mortality

was approximately two times higher among children born to mothers in North Eastern

tioned are introduced. All the regressions include an offset variable mentioned above.

Results from the negative binomial models are sometimes better ex

more convenient scale. All coefficients have been put on an exponentia

the e

er-five mortality for each unit increase in the independent variable.

Results

Depending on the variables included in the model, regional coefficien

magnitude and direction. Results in Table 2 show that the risk of child

slightly over three times higher among children born to mothers in Nya

compared to Nairobi. The risk of child mortality was highest in Nyanza Pr

9

Kenya compared to Nairobi. However, addition of control variables in the models

significantly alters the relationship between the various provinces. This indicates the

rtance of the control variables in understand child mortality risks in the

Under-five mo ality inc nce rat s for lected ind endent variables, (Kenya)

Model

relative impo

regions of Kenya.

Table 2: rt ide e ratio se ep

able/cat IV

Nairobi (CenCoastEasteNorth-Eastern NyanzNSW

1520

30-3435-40-44

cation

PrimaSe

Marital stNever Previously

Residence Rural U

Unsafe(RC) Other-safe Tap/ pipe-borne

-

*0.84 2.07

*1.27 *1.81

1.15 *1.70 *2.25

-

-

-

-

***11915.9 8

*******1*****2.38

*****

-

-

-

*0.75 *1.93

.14 *1.67

*1.07 *1.57 *2.08

****1.20 ***1.68 ***1.82 ***2.23 ***2.55

-

-

-

***11915.9 14

***

***1.70 *

*

- *1.09

-

-

-

**0.59 **1.34 **0.80

0.98

**0.74 1.02

**1.50

****1.23 ***1.71 *****2.57

-

-

-

***0.9***0.92

- ***0.73 ***0.49

***13405.1

18

***0.89

***1.85

***1.17 ***1.50

***1.92

***0.67

1.03

-

***15393.2 19

-

***0.64 ***1.24 ***0.82 ***0.87 ***1.72 ***0.75

1.00 ***1.52

-

***1.13 ***1.20 ***1.55 ***1.59 ***1.84 ***2.02

- ***0.70 ***0.43

-

***1.22 1.01

-

***1.12

- ***0.97 ***0.92

- ***0.80 ***0.63

***15997.1

23

- **1.11

**1.85 **2.26

- 7

- ***0.66***1.38

*1.08

***0.83***1.15***1.64

- ***1.12

***1.53***1.76

-

***0.40

- ***1.25

- 0.99

-

10

Vari egory I II III V Region of residence

RC) tral

al rn

a orth-Rift outh Rift est

Age group -19 (RC) -24

25-29

39

45-49 Edu

None (RC) ry

condary & over atus

married (RC) married

Currently married

(RC) rban

Water source

Toilet facility

None (RC) Other Flush

Likelihood ratio X2

Degrees of freedom

************2.55 *******

*** = p<1%; **= p<5%; *= p<10%; N=135459

Table 3: Under-five mortality incident rate ratios by region of residence and some selected indep va les enyendent riab , (K a) VINCE OF RE

North East an North Rift South Rift

8 1.90

5 4

2.

- 0

0.0.38

- 9 1.

1.09

- 1.16

- 9 1.0

1.14

- 1.00

1.23

***1.34 ***1.44

***0.75 0.** 43

***1.42 ***1.19

-*1.07

-890.

***0.85

-***0.78

***1.37 .7*1

***2.46

***0.75 .4* *0 5

***1.26 1.02

1.07

*1***1.41

.1

***0.89

*2.13

*2.*2.

39

8 0

*0.*0.

65

4 2 *

- 5 1.0

0.94

- 1.19

- 2 1.0

0.98

- ***0.74

**1.34 1.6

***1.98 ***2.17

***0.74 0.5**

- 9 4

- 0

- 5

1.31.1

***1.2

0.9***0.78

***0.88

REGION/PRO SIDENCE Variable/category

Nairobi Central Coastal Eastern Ny za West Age group

15-19 (RC) 20-24 25-29 30-34 35-39 40-44 45-49

Educational level attainment None Primary Secondary & over (RC)

Marital status Never married Previously married Currently married (RC)

Residence Rural Urban (RC)

Water source Safe Unsafe Tap/ pipe-borne (RC)

Toilet facility None Other Flush (RC) Likelihood ratio X2

Degrees of freedom Number of cases

-

1.07 0.85 1.09 0.94 1.09 1.07

-

***0.72 ***0.36

-

**1.22 0.99

- -

- 1.48 1.16

-

0.98 **0.75

***389.6

14 9476

-

1.17 1.30

***1.71 ***1.88 ***2.24 ***2.43

-

***0.64 ***0.46

-

**1.17 **0.90

-

1.12

- 0.96

***0.86

- 0.80 0.86

***877.0

15 20454

-

*1.24 ***1.46 ***1.90 ***1.97 ***2.24 ***2.40

-

***0.64 ***0.43

-

***1.22 1.05

-

1.01

- **0.92 **0.92

-

***0.85 ***0.63

***1056.6

15 12879

-

**1.49 ***1.91 ***2.45 ***2.82 ***3.35 ***3.76

-

***0.66 ***0.42

-

***1.11 **0.89

-

***1.25

- 1.02

***0.77

- ***0.76 ***0.61

***1962.8

15 24838

-

1.20 1.3

*1.77

**2.2** 3

** 5*

3

1.25

***82.9 15

2454

-

1.02 0.99

*** ***1.19

-

*

-

***

***0.63

***2053.0 15

24096

-

***1.18

** 9 ***1.90 ***2.25

-

*

-

-

-

** 2

-

***0.72

***896.7 15

20499

-

1.26 **1.48

*****2.20 ****

- **

*

***0.51

***1012.8 15

12626

-

**1.34

*** 6 ***1.68

-

* 0

***

**

-

***0.67

***683.5 15

8135 *** = p<1%; **= p<5%; *= p<10%;

11

Results in Tables 2 and 3 also provide estimates of the effect of some selected

individual- and household-level characteristics of mothers on the mortality of their

en of mothers

lity compared

kewise, these

oximately 30

percent those of mothers with a primary education. A look at region-specific result (Table

3) reveals that maternal education effects are very strong in all the individual province

a significantly

never married

usly married

mothers. This result is also reflected in the regional-level models in Table 3 except for the

North Eastern and South Rift Provinces where marital status was not significant in the

five mortality

f never married mothers and those of currently married

mo provinces, as

ong children

whose mothers were previously married.

Regression results in Tables 2 and 3 appear to be inconclusive with regard to rural-

urban child mortality differentials. Rural and urban residence coefficients were barely

significant in the models except in Coastal, Nyanza, and Western Provinces. The key

question raised is why child mortality should be lower in some rural areas of the respective

children. Results in Models IV and V (Table 2) show that in Kenya childr

with no education had approximately 60 percent higher risk of child morta

to their counterparts whose mothers had a secondary or higher education. Li

children of mothers with no education experienced mortality risks appr

models and follow the pattern discussed above.

In Table 2 results suggest that children of never married mothers had

lower mortality compared to those who were currently married. Children of

mothers experienced lower risk of mortality compared to those of previo

models at all. It also appears that there is no significant difference in under-

risks experienced by children o

thers. This fact is reflected both in Table 2 and in six out of the nine

indicated in Table 3. Under-five mortality risks were therefore highest am

12

regions and higher in others, even when the expectation is a lower mortality in the urban

areas? In fact, Model V (Table 2) suggests that overall child mortality risks in Kenya were

s. However, in Table 3 the results

men

increased the

province the

findings seem to suggest that the risk of child survival was higher among children of

mothers who resided in households using pipe-borne water compared to those who used

uns rtifact of data

oilet facilities

ept in

Central and North Eastern, where the variable type of toilet facility was not significant in the

models. The age coefficients were not significant in the model estimated for Nairobi (Table

3), however, results for other provinces generally show the pattern already observed in

Mortality risk among children increased with the age of mother. This perhaps

refl ers.

Conclusions

The purpose of this paper was to examine the regional differences in infant and child

mortality in Kenya using a count data regression procedure. The study uses the 1989 Kenya

census to examine some of the factors that are most responsible for regional differences in

under-five mortality. Kenya has in the recent past experienced substantial and sustained

significantly higher in the urban areas than in rural area

tioned earlier reveal that this is a feature of only three provinces.

In Table 2, results also show that, as expected, unsafe water sources

risk of child mortality. Considering the results in Table 3 for North Rift

afe water sources. This latter finding was unexpected, and is perhaps an a

in this province.

As expected, children of mothers from households with flush t

experienced the lowest mortality risks. This result is consistent in all provinces exc

Table 2.

ects the longer duration of exposure to mortality for children of older moth

13

decline in mortality associated with general socioeconomic development and changes

related to education, perhaps facilitated by the long political stability (Brass and Jolly, 1993;

still remain.

ity rates have

DHS suggest

998). This is

mainly attributed to the introduction of structural adjustment programs which have

substantially reduced government subsidies to the health sector and the onset of HIV/AIDS.

est under-five

gest likewise

). According

lity reflect the

importance of socioeconomic development. In this study, even after controlling for other

variables, education of mother remained significant in the regression equations. Further,

rtality rates in Nyanza and Coastal Provinces

are ese potential

mortality rates

This paper further reinforces the findings of previous studies concerning the

relationship between maternal education and child survival. The coefficients for education

were highly significant in all the models estimated, even in the regional models. The results

indicate an inverse relationship between maternal education and child mortality.

Ewbank et al., 1986). However, regional differences in under-five mortality

Evidence from the data suggests that since around 1985, under-five mortal

been increasing in Kenya. Moreover, preliminary results from the 1998 Kenya

that childhood mortality rates in Kenya increased in the recent past (NCPD, 1

In Kenya, Nyanza, Coastal, and Western Provinces registered the high

mortality risks. This finding is in line with previous studies, which sug

(Ewbank et al. 1986; United Nations, 1991; NCPD, 1998; Rutaremwa, 1999a

to Ewbank et al. (1986), the explanations for geographic differentials in morta

according to their study the high under-five mo

attributable to the prevalence of malaria and diarrhea in. Because of th

cultural, socioeconomic, environmental, and historical differences, under-five

in the different regions of Kenya have also been variable.

14

With regard to level of urbanization, the results are unclear. Theoretically, all things

being equal, living in urban areas should be associated with a higher standard of living,

esults indicate

icantly higher

fact that the

his finding is

consistent with Ewbank et al. (1986) in their Kenyan study, although their analysis of rural-

urban differentials was not definitive. Preston and Haines (1991) found that America at the

han the rural areas. They

exp nd congestion,

.

ion equations.

The variations in the relative size of the coefficient for this variable for the different regions

of Kenya indicate the regional differences in relative importance of this variable. The

sign expected since

hild mortality.

ory currently

married had the least risk of child mortality. This relationship was more pronounced for

all the provinces with exception of the North Eastern Province. The risks of child

mortality were lowest among children of mothers who were never married. What is

characteristic about this category of women (never married) is that they are young and

compared to the rest of the women they have fewer children. It is possible that the

better sanitation, and better health facilities, among other things. However, r

that for some regions of Kenya, the under-five mortality indicator was signif

for the urban areas compared to the rural areas. This is perhaps due to the

models examined variables that are remotely correlated to mortality. T

turn of the century also witnessed higher mortality in urban areas t

lain that urban areas were associated with a lot of unsanitary conditions a

which could be the case in the slum environment in some parts of urban Kenya

Type of toilet facility emerges as an important variable in the regress

ificant negative coefficient for North Rift Province of Kenya is rather un

it suggests that unsafe sources of water contribute to lowering the risk of c

This finding could perhaps be an outcome of problems related to the data.

Results with regard to marital status generally suggest that the categ

15

children of single mothers have on average more resources in terms of diet and

medication relative to children from larger families. In addition, it is likely that children

question of

osal of the child and mother may play an important role in the

surv

ortality exist

and are strong in Kenya. In addition, the factors most associated with under-five

mortality are similar in all regions of Kenya. When control variables are added in the

agnitude, and

aps regional

l attributes of

also the role

of political influence. Clearly, these factors are not fully captured by census data and

most surveys, yet the same factors shape the demographic, biological, socioeconomic,

historical, and cultural background of specific regions which in turn have an impact on

regional under-five mortality levels and patterns.

of single mothers are fostered to the extended family, where again the

resources at the disp

ival status of the child.

Finally, this study concludes that regional differences in under-five m

models, regional differences in under-five mortality tend to be reduced in m

sometimes the direction of the relationship changes altogether. Perh

differences in under-five mortality are an outcome of the underlying spatia

these regions: resource endowments including climate, soils, vegetation and

16

References

Allison, D. P. 1998a. Logit and Loglinear Analysis Using the SAS System. University of

Ank nalysis of Mortality Differentials in

Population and Employment Working

Brass, W. and C.L. Jolly (eds.). 1993. Population Dynamics of Kenya. Working Group on

ommittee on tional Academy Press.

ount Data.”

ss. Esrey, S.A. and Jean-Pierre Habicht, 1986, “Epidemiological evidence for the health

countries,” ol of Hygiene

of demographic and

ts of Mortality e Country Case Study

ed Nations. 1986.

Greenwood, M. and G. U. Yule. 1920. “An Inquiry into the nature of frequency distributions of multiple happenings, with particular reference to the occurrence

oyal

ent Variables.

National Council for Population and Development (NCPD). 1998. Kenya Demographic and

Health Survey 1997, Preliminary Results. Nairobi, Ministry of Planning and National Development.

Preston, S.H. and M.R. Haines. 1991. Fatal Years: Child Mortality in Late Nineteenth-

Century America. Princeton, NJ: Princeton University Press.

Pennsylvania.

Allison, D. P. 1998b. Categorical Data Analysis. University of Pennsylvania.

er, R. and J.C. Knowels.1977. “An Empirical AKenya at the Macro and Micro Levels.” Paper No. 60. Geneva: International Labour Organization.

Kenya: Panel on Population Dynamics of Sub-Saharan Africa, CPopulation, National Research Council. Washington, DC: Na

Cameron, A.C and P.K. Trevedi. 1998. “Regression Analysis of CEconometric Society Monographs No. 30, Cambridge University Pre

benefits from improved water and sanitation in developingEpidemiological Reviews, Volume 8, Johns Hopkins University Schoand Public Health.

Ewbank D., R. Henin, and J. Kakovole. 1986. “An integrationepidemiologic research on mortality in Kenya,” in DeterminanChange and Differentials in Developing Countries: The FivProject. New York: Unit

of multiple attacks of disease or repeated accidents,” Journal of the RStatistical Society, pp. 255-279.

Long, S. L. 1997. Regression Models for Categorical and Limited DependThousand Oaks, CA: SAGE Publications

17

18

Rut mortality: A ive study of Kenya and Uganda,” Population Studies Center, University of

Pennsylvania.

Rut mputation of missing maternity history data from the Kenya, 1989 census: A case for Nyanza Province.” Population Studies Center,

Rut rtality in less

Hill (ed.), Child Health Priorities for the 1990s. sity School of

ortality from

-36.

Research”. In , A Lam, A.I.

rk, Oxford: Oxford University Press.

ic Differentials, lopment. New

Sales No. E.91.XIII.13. Weinreb, A.A. 2000. “Integrating respondents, community, and the state in the analysis

of contraceptive use in Kenya.” Ph.D. dissertation in Demography and Sociology, University of Pennsylvania.

aremwa, Gideon. 1999a. “Regional differences in infant and child comparat

aremwa, Gideon. 1999b. “Multiple i

University of Pennsylvania (unpublished).

stein, S.O. 1992. “Levels, trends and differentials in infant and child modeveloped countries,” in Kenneth Report of Seminar held June 20-22, 1991 at Johns Hopkins UniverHygiene and Public Health, Baltimore, MD.

Trussell, J. and S.H. Preston. 1982. "Estimating the covariates of childhood mretrospective reports of mothers," Health Policy and Education 3(1): 1

Trussell, J. and G. Rodriguez. 1990. “Heterogeneity in Demographic Convergent Issues in Genetics and Demography. Eds. Adams, J.D.Hermalin and P.E. Smouse. New Yo

__________. 1991. Child Mortality in Developing Countries: SocioeconomTrends and Implications. Department of Economic and Social DeveYork: United Nations,

Last Working Papers published

ACAP W.P. # 1: Tukufu Zuberi and Ansom Sibanda, Fertility Differentials in sub-

Saharan Africa: Applying Own-Children Methods to African Censuses, January 1999.

ACAP W.P. # 2: Herbert B. S. Kandeh, Using indigenous knowledge in the demarcation

of the enumeration areas: A case study of Banta Chiefdom, Moyamba District, Sierra Leone, January 1999.

ACAP W.P. # 3: Etienne Van de Walle, Households in Botswana: An exploration,

February 1999. ACAP W.P. # 4: Amadou Noumbissi, Mortality analysis using Cameroon 1987 census

micro data, March 1999. ACAP W.P. # 5: Monde Makiwane Fertility in rural South Africa: The case of Transkei,

March 1999.

ACAP W.P. # 6: Tukufu Zuberi and Akil K. Khalfani, Racial Classification and Colonial Population Enumeration in South Africa, March 1999.

ACAP W.P. # 7: Tukufu Zuberi and Akil K. Khalfani, Racial Classification and The

Census in South Africa, 1911-1996, March 1999.

ACAP W.P. # 8: Amson Sibanda and Tukufu Zuberi, Contempory Fertility Levels and Trends in South Africa: Evidence from Reconstructed Birth Histories, April 1999.

ACAP W.P. # 9: Etienne van de Walle. Where are the Children of Botswana? June 1999

ACAP W.P. # 10: Gideon Rutaremwa. Regional Differences in Infant and Child Mortality: A comparative Study of Kenya and Uganda, July 1999.

ACAP W.P. # 11: Ayaga A. Bawah and Tukufu Zuberi. Estimating Childhood Mortality from Census Data in Africa: The case of Zambia, August 1999.

ACAP W.P. # 12: Amadou Noumbissi; Tukufu Zuberi and Ayaga A. Bawah. Multivariate Analysis Using Grouped Census Data: An Illustration on Estimating the Covariates of Childhood Mortality, October 1999.

ACAP W.P. # 13: Pierre Ngom; Ibrahima Sarr and Alieu Gaye. Ethnic Diversity and Assimilation in Senegal: Evidence from the 1988 Census. February 2000.