Family-level clustering of childhood mortality risk in Kenya

1

Family-level clustering of
childhoodmortality risk in Kenya

• D. Walter Rasugu Omariba
• Department of Sociology
• Population Studies Centre
• University of Western Ontario
• London, Ontario

2
Background

• Mortality decline in Kenya began in late 1940s.
• E.g. under-five mortality 220 in 1958-62 period,
  declined to 89 in 1984-1989 period
• Reversals in the downward trend started in 1986
  (see figure 1).
• Infant mortality increased by 24 and
• Under-five mortality by 25 in 1988-98 period.

3
Figure 1 Child mortality trends 1974-1998, Kenya

• Source National Council for Population and
  Development and Macro International, 1989, 1994
  1999.

4
Existing research

• Focuses on determinants and differentials of
  mortality (See, for instance, Kibet, 1981 Ewbank
  et al., 1986 Kichamu, 1986 Omariba, 1993
  Obungu et al., 1994 Ikamari, 2000).
• This studys focus
• Familial child death clustering
• In the literature, defined in two ways
• 1) Expected vs. observed- Higher observed deaths
  indicate death clustering
• 2) Control for unobserved heterogeneity through
  inclusion of random effects in models-
  correlation of risks at different levels.

5
Rationale

• Random-effects models used yet to be applied on
  Kenyan data.
• Child mortality remains an important public
  health issue.
• Reducing mortality important for sustaining
  countrys incipient fertility transition.

6
Sources of unobserved heterogeneity

• Differential competence in childcare (Das Gupta,
  1997).
• Biological factors e.g. genetically determined
  frailty, improvident maternity syndrome (Guo,
  1993 Das Gupta, 1997).
• Socioeconomic, cultural factors and environmental
  factors.
• All unmeasured and unmeasurable factors.

7
Death clustering?

• In this study
• Measured by unobserved heterogeneity term
  indicating correlation of risks in family.
• Most studies only select one child, truncate data
  by certain date or ignore first child- Biased
  results especially when variables such as
  preceding birth interval and survival status are
  considered.

8
Implications of data structure

• Children in same family are more alike than
  children from different families.
• covariates estimates biased.
• Consequences of violation of independence
• standard errors of parameters underestimated
  spurious precision.
• biases baseline hazard duration pattern downward
  in survival analysis.

9
Implications of data structure

• Random-effects models Correct for the biases in
  parameter estimates, provides correct standard
  errors and correct confidence intervals and
  significance tests
• Separates impact of individual and social context
• If contextual effects significant, using a random
  effect (or multilevel model is reasonable). If
  not, then we need only adjust the error term for
  dependence of units.

10
Data and methods

• Data source Demographic and Health Survey for
  Kenya, 1998.
• 7,881 women 15-49, all marital statuses from
  8,380 households and 8,233 eligible women.
• 3,407 husbands/partners of the women
• Largely rural sample, 81.4 of the womens sample
• Methods
• Weibull hazard models and random-effect hazard
  models.
• The latter tests for family-level variance.

11
Conceptual framework

• Study is guided by the Mosley and Chen (1984)
  proximate determinants model (see Figure 2).
• Individual characteristics Migration status,
  education, year of birth, ethnicity, religion,
  survival status of preceding child, birth
  interval, birth order and maternal age at birth.
• Household characteristics socioeconomic status,
  sanitation and source of water.

12
Figure 2 Conceptual framework for studying
the determinants of infant and childhood
mortality
Proximate Determinants
Distant Factors
-Reproductive healthcare behaviour e.g. prenatal
care, place of delivery, delivery care, tetanus
injection, breastfeeding -Biodemographic
factors e.g. maternal age at birth, birth
interval, birth order, age at marriage, child
loss experience -Household environmental
conditions e.g. source of water, toilet
facility.
-Socio-economic factors e.g. maternal paternal
education, place of residence, region, migration,
occupation, household socioeconomic status,
marital status, year of birth, period of child
birth. -Socio-cultural factors e.g. religion,
ethnicity.
Outcome Variable
Risk of child death
13
Data description

• Of the 7881, 5716 had at least one child, while
  2165 had never had a child.
• 23348 children born to 5716 women (family)
• 2325 children had died before their fifth
  birthday
• Infancy- 1620(0-12 months)
• Childhood- 705 (Age 13-59 months)

14
Table 3 Distribution of children and child
deaths per family in Kenya, DHS 1998
Children per/fam     Deaths in family Deaths in family Deaths in family         Percent of   Percent of  
  0 1 2 3 4 5 6 7 8 Total Children Deaths
1 1012 87 0 0 0 0 0 0 0 1099 4.7 3.7
2 884 99 8 0 0 0 0 0 0 991 8.5 4.9
3 632 130 16 0 0 0 0 0 0 778 10.0 7.0
4 523 131 30 3 2 0 0 0 0 689 11.8 9.0
5 366 128 36 11 1 0 0 0 0 542 11.6 10.2
6 327 115 47 15 3 2 0 0 0 509 13.1 11.9
7 193 100 42 14 9 1 0 0 0 359 10.8 11.5
8 129 81 35 19 7 4 0 0 0 275 9.4 11.0
9 105 62 29 18 9 3 0 0 1 227 8.7 10.0
10 41 40 23 18 8 6 2 1 0 139 5.9 9.5
11 14 11 12 6 5 2 3 0 0 53 2.5 4.3
12 6 6 6 3 12 2 1 2 0 38 2.0 4.5
13 1 2 1 4 2 0 2 0 0 12 0.7 1.6
14 0 1 0 0 0 1 1 0 0 3 0.2 0.5
15 0 0 1 0 0 0 0 1 0 2 0.1 0.4
Total 4233 993 286 111 58 21 9 4 1 5716 100 100
of children 62 22 8 4 2 .8 .5 .2 .03 100 ----   -----
of deaths 0 43 25 14 10 5 2 1 0.3 100    
15
Does clustering exists?

• Over 80 percent of the children belong to
  families contributing two or more children to the
  sample.
• Families with six or more children comprise about
  28 percent of the families yet contribute over
  half of the children.
• 57 percent of the deaths occurred to 8.6 percent
  of the families with two or more deaths.
• About 2 percent of the families contribute four
  or more deaths together accounting for about 18
  percent of the deaths.

16
Results

• There is significant unobserved heterogeneity
  both in infancy and childhood (Tables 3 4)
• The estimated random parameters, ?, in the models
  with unobserved heterogeneity are 0.40 and 0.78
  for infant and child mortality respectively.
• There is significant familial variation in the
  risk of infant and child death.
• Maternal education, period of birth, ethnicity,
  type of toilet facility, birth interval and
  maternal age at birth of child important for both
  infant and child survival (Tables 12).
• Migration status, religion, survival status of
  previous child and birth order significant only
  for infant mortality, while household SES
  significant only for child mortality.

17
Results

• There are large ethnic differences in risk of
  death with children Luo mothers being most
  disadvantaged.
• Secondary or higher education associated with a
  22 and 42 reduction in risk of infant
  mortality and child mortality respectively.
• Risk of infant death higher for children born
  after 1990, while that of child death is higher
  for all children born after 1985.
• The risk of infant death is higher for children
  whose sibling died, were born less than 19 months
  after preceding sibling, and when the mother was
  less than 20.

18
Conclusions

• The determinants of death have different effects
  on infant and childhood mortality. Biodemographic
  factors have greater effect in infancy, while
  education and ethnicity have greater effect in
  childhood.
• Suggests varied policy actions
• Infancy longer birth intervals through family
  planning and breastfeeding, later age at birth
  etc.
• Childhood improvement in education,
  socioeconomic status and poverty eradication
  programs.

19
Conclusions

• Death clustering is non-ignorable Needs further
  research
• Healthcare factors- Information available only
  for children born three years before the survey.
• Qualitative research at community level.
• Death clustering, another measurement Consider
  unobserved heterogeneity in the context of each
  womans sequence of births. The heterogeneity
  term used in this paper does not reflect this
  fact.