IJE vol.32 no.6 © International Epidemiological Association 2003; all rights reserved.
International Journal of Epidemiology 2003;32:1041–1051
DOI: 10.1093/ije/dyg241
THEORY AND METHODS
Predicting the distribution of under-five
deaths by cause in countries without adequate
vital registration systems
Saul S Morris,1 Robert E Black2 and Lana Tomaskovic3
Accepted
14 May 2003
Background The absence of complete vital registration and atypical nature of the locations
where epidemiological studies of cause of death in children are conducted make
it difficult to know the true distribution of child deaths by cause in developing
countries. A credible method is needed for generating valid estimates of this
distribution for countries without adequate vital registration systems.
Methods
A systematic review was undertaken of all studies published since 1980 reporting
under-5 mortality by cause. Causes of death were standardized across studies,
and information was collected on the characteristics of each study and its population. A meta-regression model was used to relate these characteristics to the
various proportional mortality outcomes, and predict the distribution in national
populations of known characteristics. In all, 46 studies met the inclusion criteria.
Results
Proportional mortality outcomes were significantly associated with region,
mortality level, and exposure to malaria; coverage of measles vaccination, safe
delivery care, and safe water; study year, age of children under surveillance, and
method used to establish definitive cause of death. In sub-Saharan Africa and in
South Asia, the predicted distribution of deaths by cause was: pneumonia (23%
and 23%), malaria (24% and 1%), diarrhoea (22% and 23%), ‘neonatal and
other’ (29% and 52%), measles (2% and 1%).
Conclusions For countries without adequate vital registration, it is possible to estimate the
proportional distribution of child deaths by cause by exploiting systematic
associations between this distribution and the characteristics of the populations in
which it has been studied, controlling for design features of the studies themselves.
Keywords
Cause of death, mortality, preschool child, infant mortality, Sub-Saharan Africa Asia
Setting appropriate public health targets and measuring
progress toward their achievement requires information on
patterns of disease and factors that increase or decrease risk.
Unfortunately, the epidemiological evidence base for the
distribution of child mortality by cause is inadequate to support
1 Department of Epidemiology and Population Health, London School of
Hygiene & Tropical Medicine, London, UK.
2 Bloomberg School of Public Health, Johns Hopkins University, Baltimore,
MD, USA.
3 Rue Ancienne, 1227 Carouge, Geneva, Switzerland.
Correspondence: Saul S Morris, Public Health Nutrition Unit, London School
of Hygiene & Tropical Medicine, 49–51 Bedford Square, London WC1B 3DP,
UK. E-mail: [email protected]
sound public health decision-making in most countries or even
at regional and global levels. Both of its main sources of data are
flawed. Vital registration data, the first source, are usually
collected in countries with under-5 mortality rates that are low
by global standards. For example, in the whole of Africa, only
the island states of Mauritius and the Seychelles have vital
registration systems with coverage of 95% or more.1 South
Africa has the most complete system in continental SubSaharan Africa. Although an analysis of the quality of this
system in the early 1990s found a coverage of only 19% of all
births,2 it is said to have improved since that time. India has a
sample registration system, the quality of which has also been
questioned.3 The second source of relevant data consists of
epidemiological studies of cause-specific child mortality in
1041
1042
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
specific populations. These studies are limited in number, and
are generally conducted in populations that are either atypically
easy to access or have atypically high mortality rates. The
resulting data have limited utility for public health planning,
even in the countries where the studies are conducted, unless a
valid means can be found of extrapolating to other populations
with—most likely—different characteristics.
This paper outlines a method for extracting generalizable results
on the proportional distribution of under-5 deaths by cause in
countries without reliable vital registration systems. This method
involves systematic searching for all relevant epidemiological
studies, standardizing the way causes of death are categorized
across studies, and using regression techniques to relate the
proportional distribution of deaths by cause in each study to a
limited number of characteristics of the study populations and
study designs. Finally, these relationships are exploited to estimate
the proportional distribution of deaths by in national populations
of known size and characteristics. We present estimates of the
proportional distribution of under-5 deaths by cause in two
regions that together accounted for 75% of all child deaths in the
year 2000: sub-Saharan Africa and South Asia.
Materials and Methods
We conducted a systematic search of all studies published since
1980 that report the distribution of child mortality by cause.
Both MEDLINE and POPLINE were searched using the terms
(infant OR child) AND (death* OR mortality OR fatal outcome)
AND (ARI OR ALRI OR pneumonia OR respiratory) AND
(diarrhoea OR diarrhoea). No restriction was placed on
publication language. The reference sections of these studies
were then reviewed to identify additional studies. We also
contacted researchers conducting cause-specific mortality
reviews for the Child Health Epidemiology Reference Group of
the World Health Organization to ensure that any appropriate
articles included in these reviews were also included in the
current study. Members of the Reference Group, as well as
members of the staff of both the London School of Hygiene &
Tropical Medicine and the Bloomberg School of Public Health at
Johns Hopkins University, were asked to identify any
unpublished studies known to them.
The initial pool of studies was then screened to ensure that
they met the inclusion criteria for the present study. These were
(1) surveillance period no earlier than 1980, (2) mortality
surveillance for either an exact multiple of 12 months (prospective or retrospective surveys) or exactly the same age range for all
study subjects in the case of birth cohorts, to minimize seasonal
effects, (3) use of a standardized verbal autopsy questionnaire
with or without supporting clinical records to ascertain cause of
death, (4) a unique cause (or fully identified combination of
causes) for each death, and (5) no more than one-third of all
deaths attributed to ‘undetermined’ causes. Forty-seven studies
were identified that met these criteria, 46 of them described in
43 different publications, and one as yet unpublished in any
form (4–44; Henry Perry, Hospital Albert Schweitzer, personal
communication, 2002).
A total of 231 different ‘causes’ of death were identified in the
data set. Causes were grouped into 14 subsets: diarrhoea,
pneumonia, malaria, measles, diarrhoea plus pneumonia
(combined), diarrhoea plus malaria (combined), pneumonia plus
malaria (combined), malnutrition, ‘neonatal causes’, injury, sepsis/
septicaemia/meningitis, other specified, other unspecified, and
undetermined. We contacted study investigators and asked them
to identify the cause of death when non-specific terms were used
in the original publication. Where needed we also asked them to
distinguish between measles deaths and deaths from other
vaccine-preventable causes.
Most studies identified only a single cause for each death.
Deaths attributed to combined causes such as diarrhoea plus
pneumonia were therefore re-distributed between their singlecause components. For example, if in study j there were d1
deaths attributable to diarrhoea, d2 attributed to pneumonia,
and d1.2 attributed to the combined effect of the two illnesses,
then the d1.2 combined-cause deaths were redistributed to
diarrhoea and pneumonia in the ratio d1:d2.
Deaths attributed to malnutrition were re-allocated among all
other categories of infectious illness based on the relative
importance of the single-cause deaths in the same studies.
Labelling a small proportion of deaths as ‘due to malnutrition’
would underestimate the true contribution of malnutrition to
child mortality in poor communities. Underweight is a risk
factor for deaths from all infectious causes, with attributable
fractions of 52% for pneumonia, 61% for diarrhoea, 57% for
malaria, and 45% for measles.45
The nature of the available data thwarted our intention to
examine the proportion of deaths occurring in the neonatal
period. Ten of the 47 studies did not report neonatal deaths
separately, choosing instead to include them in a residual
category. Even studies that did separate neonatal deaths used
one of two irreconcilable approaches to categorize them: either
age at death or—more commonly—disease category (tetanus,
congenital malformation, etc.). We therefore combined deaths
from ‘neonatal causes’ with the remaining deaths from injury,
sepsis/septicaemia/meningitis, other specified causes, and other
unspecified causes into a single residual category, which we call
‘neonatal and other causes’.
We collected information on the characteristics of each study
and its population, including: study location (rural/urban, UN
region, longitude, latitude, and altitude); range of years encompassed by the mortality surveillance; age of children under
surveillance; overall level of mortality encountered; access to safe
drinking water (using a standardized definition); vaccination
coverage rates; percentage of births attended by qualified
personnel; population at risk of falciparum malaria; adult female
literacy rate; and the anthropometric status of children in the
study population. All relevant information in the publications
themselves was abstracted onto a standard record form. Partially
completed forms were then sent to investigators with a request
that they provide missing information. For data not provided by
investigators, we identified other studies conducted in the same
or similar sites within a few years of the index study. In the
absence of more specific information, estimates were taken from
national or sub-national surveys or population censuses
conducted within a few years of the index study. Estimates of the
proportion of the population at risk of falciparum malaria were
taken from Mara LITe, version 3.0.0 (Mapping Malaria in Africa,
Durban, South Africa) for African populations, and were supplied
by the Centers for Disease Control Division of Parasitic Diseases
(Rick Steketee, personal communication, 2002) for all other
populations.
PREDICTING DEATH IN COUNTRIES WITHOUT VITAL REGISTRATION
Estimates of the level of all-cause mortality were transformed
to a single metric (the risk of a child dying before reaching age
5, 5q0), because they were variously reported as rates, risks, or
ratios in the publications and often for non-standard age
groups. We generated this standard measure by (1) choosing an
appropriate model life table, usually from the series developed
by the INDEPTH demographic surveillance sites,46 (2) using the
method of Brass and Blacker47 to ‘fill out’ the model life table
to every month of age between birth and 10 years, (3) varying
the level of mortality in the model life table on the logit scale48
until the measure of mortality reported in the publication was
replicated in the model, and (4) reading off the implied level of
5q0. Two studies did not report any information that would
permit the estimation of 5q0. For one we developed an estimate
based on sub-national data available from a survey conducted at
approximately the same time; the other study was dropped
from the analysis.
We used a regression framework to relate the distribution of
mortality from different causes to the characteristics of the study
designs and study populations. One approach would have been to
use the proportional mortality, P, due to cause k (k = diarrhoea,
pneumonia, malaria, measles, other, undetermined) as the dependent variable in six separately estimated equations. However, this
strategy ignores the fact that for any given study, the six proportions must sum to one. The correct regression model therefore
involves just five equations plus a constraint forcing the
proportions to sum to one.
Salomon & Murray49 considered this problem in the context of
overall mortality from communicable and non-communicable
causes and injury. Although they adapted their model from the
work of Katz and King,50 the original mathematical development
is attributed to Aitchison.51 In brief, five dependent variables are
developed by dividing the proportion of deaths in each study due
to cause k by the proportion due to a selected ‘base’ cause K, and
taking the natural logarithm of this value. The vector of
dependent variables is then assumed to be multivariate Normal,
and a linear function of the explanatory variables in the model.
The model can be written as
ln(Pjk/PjK) = α + β STUDY_CHARACTERISTICSj
+ γ POPULATION_CHARACTERISTICSj + jk
where alpha, beta, and gamma are regression coefficients, and
epsilon is a random error term specific to the study and the
particular cause k. The resulting set of five equations could
be solved by Ordinary Least Squares regression. However, to
account for the correlated nature of the five equations, we
estimated the entire set simultaneously using Seemingly
Unrelated Regression.52
The compositional model outlined above is invariant to the
choice of ‘base’ cause, K, holding constant all other variables in
the model. We experimented first with taking diarrhoea as our
base cause, and then pneumonia, because both causes of death
were present in each of the 46 studies under analysis. Presumably
because many variables are associated equally with mortality
from diarrhoea and mortality from other causes, it was more difficult to identify significant associations with the k-cause:diarrhoea
ratios than with the k-cause:pneumonia ratios. The latter
specification is therefore presented in this paper.
We experimented with various approaches to weighting the
different studies in the regression analysis, based on adaptations
1043
of the methods described by Thompson & Sharp.53 ‘Optimal’
weights were almost identical for the studies with the smallest
and largest sample sizes, and the results presented in this paper
are therefore unweighted. To develop the final model, we first
used Ordinary Least Squares regression to explore bivariate
associations one cause (or rather, ratio of causes) at a time. At
an early stage we discarded child anthropometric status, urban/
rural location, and adult female literacy as potential explanatory
variables because they were not associated with any of the
dependent variables (ratios) after adjusting for 5q0. For the
remaining variables, possible non-linearities were explored
ratio by ratio and explanatory variable by explanatory variable
using fractional polynomial models,54 a highly flexible method
of curve fitting. We examined statistical best-fit models visually
to see whether they could have been influenced by outliers, and
made minor adjustments in a few cases. All associations
significant at the P 0.2 level were considered for entry in
multivariate models. We first selected a parsimonious set of
explanatory variables for each dependent variable using
forward stepwise regression, and then re-ran the same models
using Seemingly Unrelated Regression. Only variables
significant at the P 0.1 level were retained in the final model.
This relatively generous cut-off was selected because of the
small sample size and associated limited power of the analysis.
For each model we report the per cent of variance explained by
the predictors. We denote this ‘R2’, with the single quotes
cautioning that R-squared is no more than descriptive when
generalized least squares estimators are used, as in this case.55
Because of missing values for one or more covariates, 38 studies
were included in the final model.
Finally, the regression coefficients from the final model were
used to generate predictions of the distributions of under-5
deaths by cause for the year 2000 for countries in sub-Saharan
Africa and South Asia (Afghanistan, Bangladesh, Bhutan, Nepal,
Pakistan, and Sri Lanka). The predictions were based on the
characteristics and populations of these countries in the year
2000, with the prediction variable ‘surveillance year’ also set to
2000. Each state in India was treated as a separate country
because of their size and the relative ease of obtaining necessary
data. Because the model-development data set only included
populations with under-5 mortality rates of 26 per 1000, no
predictions were made for countries/states in these regions with
lower mortality rates (Seychelles and Mauritius in Sub-Saharan
Africa, and Sri Lanka and Kerala state in South Asia). Population
data and mortality rates were provided by WHO/EIP (Colin
Mathers, personal communication, 2002) with one exception.
For the Indian states, population data were taken from the 2001
census and mortality rates were taken from the 1998–1999
National Family Health Survey, which is representative at the
state level. Coverage data for measles-containing vaccine were
official WHO/UNICEF/World Bank estimates,56 except for the
Indian states, for which values were taken from the 1998–1999
survey. Estimates of covariate values were taken from the
UNICEF statistics website (www.childinfo.org) and for India,
from the survey. The regional estimates presented here do not
include countries or Indian states with incomplete data for the
predictor variables. Countries excluded on this basis are the
Comoros, Congo, Djibouti, Gabon, and Liberia in sub-Saharan
Africa, and Afghanistan, Tripura, and six small Indian union
territories in South Asia. It should be noted that since the
1044
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
estimation model implied that different methods of determining
the definitive cause of death resulted in different proportional
distributions, it was necessary to choose a uniform method for
the predictions; we specified this to be the use of a standardized
diagnostic algorithm, a method which minimizes the proportion
of deaths attributed to ‘unknown’ causes.
It was not possible to validate the model externally using vital
registration data because no country in sub-Saharan Africa or
South Asia with an under-5 mortality rate of 26 per 1000 has
complete or near-complete vital registration. We therefore
opted to validate the model internally using jackknife
estimation.57 This involves re-estimating the model m times
(where m is the total number of studies contributing to the
regression model), each time omitting one study. On each run,
the estimated regression coefficients are used to predict the
proportional mortality outcomes in the omitted study. It was
thus possible to determine the difference between the actual
and predicted proportional mortality outcomes for each study
in the regression sample. The mean difference between
predicted and actual values can be interpreted as a measure
of bias, and the standard deviation of the differences can be
interpreted as the standard error of an out-of-sample prediction
based on all the data. Assuming these standard errors are
distributed Normally and independently for all countries in the
same region, we then estimated 95% CI for the regional
aggregates (proportional mortality by cause for the whole of
sub-Saharan Africa or the whole of South Asia) using Monte
Carlo-type simulations. This was done by adding a random
disturbance (of mean zero and standard deviation equal to the
estimated prediction standard error) to the predictions for each
country and then aggregating to the regional level; the whole
procedure was repeated 1000 times, and 95% CI for each
proportion were estimated by determining the 2.5th and 97.5th
centiles of the relevant distributions.
All analyses were undertaken using Stata 7.0 (Stata Corporation,
College Station, TX).
Results
Among the 46 studies for which 5q0 could be estimated, 19
(41%) were conducted in South Asia, with 8 studies each from
Bangladesh and India. A further 18/46 studies (39%) were
conducted in sub-Saharan Africa, with 12 countries represented.
No studies were identified from Central America, Oceania, or
Central or Western Asia, and the three studies from Northern
Africa were all from a single country (Egypt). Table 1 shows, for
each of the 46 studies, the proportional distribution of mortality
by cause after recoding and re-allocation of deaths attributed to
malnutrition and/or combined causes.
Table 2 presents selected characteristics of the study populations. The studies have been divided into five equal-sized
groups, based on the under-5 mortality risk in each population.
The African studies were disproportionately concentrated in
the highest mortality quintiles, while the studies from South
Asia were concentrated in the middle three quintiles. All of
the studies from the Western hemisphere were in the lowest
mortality quintile. The studies were distributed around an
average mid-study surveillance year of 1990, with lowmortality studies more recent than high mortality studies,
reflecting the secular downward trend in child mortality. Low
mortality levels were associated with higher proportions of
deaths due to ‘neonatal and other’ causes, and lower proportions due to measles and—especially—malaria.
Thirty-nine studies included study children in mortality
surveillance from birth, while eight initiated surveillance at
older ages, up to age one year. Six studies were true cohort
studies, while 30 identified a relevant population crosssectionally and then followed them through time. Eleven were
cross-sectional surveys with recall of deaths over a specified
period of time. Fourteen studies required multiple experts to
reach consensus on the cause of death, while nine required only
a majority decision. Four further studies used a standardized
diagnostic algorithm and three a computerized algorithm to
assign cause of death, while five studies relied on a single expert
without a standardized algorithm. Twelve studies did not specify
the method they used to assign cause of death.
Table 3 shows the results of the meta-regression analysis, based
on 38 studies with no missing covariate information. The
regression coefficients were difficult to interpret because the
dependent variables were logarithms of the ratios of two
proportions, and the model includes several complex nonlinearities. The Table therefore presents ‘interpretations of effects’
rather than the coefficients themselves. A table of coefficients
and their standard errors is available from the corresponding
author.
The regression results show that exposure to endemic,
African-pattern falciparum malaria was associated with malaria
deaths, and that measles vaccination was protective against
measles deaths. Other effects include a greater proportion of
deaths from ‘neonatal and other’ causes (relative to pneumonia
deaths) at the lowest levels of overall under-5 mortality, a lower
proportion of deaths due to pneumonia (relative to malaria
deaths) associated with better coverage of safe delivery care,
and a secular decline in the ratio of measles deaths to
pneumonia deaths. The model results also show that design
features of mortality studies can affect the observed proportions
of deaths due to different causes: the way that the definitive
cause of death was determined was associated both with the
proportion of deaths declared underdetermined and with the
balance between diarrhoea and ‘neonatal and other’ cause
deaths relative to pneumonia deaths. Excluding the youngest
infants from mortality surveillance was also found to depress
artificially the proportion of deaths attributable to pneumonia
(relative to diarrhoea deaths). Whilst virtually all the betweenstudy variability in the ratio of malaria deaths to pneumonia
deaths (‘R2’ = 0.90) could be explained as a function of the
explanatory variables considered, as could most of the betweenstudy variability in the ratio of deaths from ‘other causes’ to
pneumonia deaths (‘R2’ = 0.59), the model explained less than
half the variability in the other ratios.
The proportional distribution of deaths by cause was
estimated for 41 countries in sub-Saharan Africa, accounting for
99% of all child deaths in the region in the year 2000. The
proportions were converted to numbers of deaths by cause by
multiplying them by the estimated total number of under-5
deaths in each country in 2000. The resulting proportional
distribution of deaths by cause for the whole region is shown in
Figure 1. The four major causes of death—‘neonatal and other’
causes, pneumonia, malaria, and diarrhoea—were each found
to be responsible for between 20% and 30% of all child deaths.
Table 1 Data used in the meta-regression analysis
Total number
of under-5
Reference Country
deaths
Proportional distribution by cause
Diarrhoea Pneumonia
Malaria
Measles
Neonatal and
other causes
Undetermined
4
Ghana
15
6.7%
33.3%
40.0%
0.0%
20.0%
0.0%
5
India
221
19.2%
28.1%
0.0%
0.0%
43.2%
9.5%
5
India
606
26.1%
27.7%
0.0%
2.5%
43.7%
0.0%
6
Bangladesh
180
16.7%
25.0%
0.0%
0.0%
58.3%
0.0%
7
India
73
18.8%
20.2%
21.6%
11.5%
27.8%
0.0%
8
India
297
13.2%
27.7%
2.0%
0.8%
55.9%
0.3%
9
Bangladesh
828
16.3%
32.7%
0.0%
4.0%
31.6%
15.5%
10
Bangladesh
678
17.8%
29.5%
0.0%
3.6%
34.2%
14.9%
11
Brazil
215
43.0%
11.3%
0.0%
0.0%
33.6%
12.1%
12,13
Brazil
244
11.9%
13.5%
0.0%
0.0%
63.5%
11.1%
14
Egypt
629
28.1%
19.4%
0.0%
2.4%
50.1%
0.0%
14
Egypt
774
31.8%
17.3%
0.0%
2.5%
48.4%
0.0%
15
Bangladesh
1852
9.9%
23.7%
0.0%
2.5%
53.1%
10.8%
16
Ghana
451
15.3%
12.4%
36.1%
0.0%
20.2%
16.0%
17
The Gambia
292
17.5%
11.6%
34.9%
1.4%
12.3%
22.3%
18
The Gambia
856
18.0%
22.3%
18.5%
0.4%
13.9%
27.0%
20
Burundi
160
19.7%
14.6%
34.9%
8.9%
19.5%
2.5%
21
Senegal
1158
23.5%
12.6%
25.1%
0.3%
31.1%
7.3%
Excluded due
to missing
covariates
X
X
22
Nigeria
314
14.4%
9.8%
43.5%
8.7%
21.9%
1.6%
X
23
Bangladesh
1354
51.1%
11.5%
2.4%
10.5%
20.2%
4.3%
X
24
Ghana
495
29.7%
9.1%
17.2%
14.7%
7.9%
21.4%
25
The Gambia
184
16.3%
17.4%
13.6%
3.3%
39.1%
10.3%
26
India
286
15.0%
10.5%
0.0%
2.1%
58.0%
14.3%
27
China
3075
10.0%
35.5%
0.0%
2.6%
52.0%
0.0%
28
Somalia
88
27.5%
27.5%
20.7%
0.0%
24.2%
0.0%
29
The Gambia
3776
9.3%
19.6%
26.0%
0.3%
19.5%
25.2%
30
South Africa
156
38.9%
9.7%
1.8%
2.7%
46.9%
0.0%
31
India
64
30.4%
23.2%
0.0%
5.4%
41.1%
0.0%
32
Pakistan
46
23.9%
39.1%
8.7%
6.5%
21.7%
0.0%
33
Pakistan
172
39.5%
12.2%
0.0%
0.0%
48.3%
0.0%
34
Brazil
111
8.1%
6.3%
0.0%
0.0%
79.3%
6.3%
35
Guinea-Bissau
153
32.2%
11.0%
15.1%
5.5%
22.0%
14.4%
36
Bangladesh
217
18.2%
19.9%
0.0%
0.0%
59.5%
2.3%
36
Bangladesh
261
19.4%
28.3%
0.0%
0.0%
47.0%
5.4%
37
Tanzania
596
14.8%
28.2%
29.1%
1.1%
20.3%
6.5%
38
Indonesia
139
7.2%
42.4%
0.0%
2.2%
48.2%
0.0%
39
India
80
41.8%
3.8%
0.0%
15.2%
39.2%
0.0%
40
Guinea
330
10.4%
23.9%
32.0%
2.0%
29.6%
2.1%
41
Ethiopia
435
8.7%
19.3%
0.0%
14.5%
57.5%
0.0%
42
Nepal
440
24.8%
19.1%
0.0%
5.1%
23.5%
27.5%
43
Ghana
42
21.4%
9.5%
11.9%
4.8%
52.4%
0.0%
43
India
59
33.9%
20.3%
0.0%
0.0%
45.8%
0.0%
0.0%
43
Peru
8
25.0%
12.5%
0.0%
0.0%
62.5%
44
Egypt
198
42.4%
28.8%
0.0%
0.0%
25.8%
3.0%
Bangladesh
283
26.0%
26.9%
0.0%
3.3%
37.1%
6.7%
99
11.5%
28.9%
0.0%
0.0%
29.3%
30.3%
Personal
communication
Haiti
X
X
X
X
1046
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
Table 2 Characteristics of the 46 study populations
Quintiles of under-5 mortality risk
I
II
III
IV
10
9
9
9
9
26.0–77.9
78.3–112.8
113.0–154.2
154.3–174.9
178.2–303.0
South-central Asia
3
5
5
5
1
Sub-Saharan Africa
2
1
3
4
7
South America/Caribbean
5
No. of studies
Under-5 mortality risk per 1000, range
V
UN region
South-eastern or Eastern Asia
1
Northern Africa
2
1
1992.7
1992.1
1991.0
1989.6
1988.6
N
10
9
8
9
7
%
62.0%
80.6%
56.6%
60.5%
56.7%
N
10
9
9
7
7
%
71.6%
22.9%
31.7%
13.7%
30.6%
(Mid-)study year, mean
Access to safe water
Per cent births attended by a qualified professional
Measles vaccine coverage rate
1
N
10
8
8
8
7
%
79.1%
82.3%
61.2%
59.0%
67.9%
Diarrhoea
25.1%
20.6%
23.5%
19.6%
20.1%
Pneumonia
12.6%
25.5%
21.8%
23.4%
19.4%
14.3%
Distribution of deaths by cause, mean per cent
Malaria
1.4%
2.9%
14.4%
15.3%
Measles
2.5%
1.4%
4.0%
3.0%
5.6%
51.1%
44.1%
29.3%
29.0%
34.3%
7.4%
5.5%
7.1%
9.7%
6.3%
Other
Undetermined
21.9%
23.0%
28.9%
0.1%
1.4%
0.1%
52.2%
Diarrhoea
Malaria
Measles
Undetermined
Pneumonia
Neonatal and other
23.7%
23.2%
23.3%
1.9%
0.2%
Figure 1 Distribution of under-5 deaths by cause, 2000. AIDS deaths are not accounted for, due to lack of studies in the affected areas
The proportional distribution of deaths by cause was also
estimated for South Asia, for areas accounting for 93% of child
deaths in the region in 2000. The predicted distribution of
deaths by cause (Figure 1) differed from that of sub-Saharan
Africa, with over 50% of child deaths being due to ‘neonatal
and other’ causes. Equal proportions of deaths were attributed
to diarrhoea and to pneumonia. Few measles deaths were
predicted except in several Indian states with low vaccination
coverage, and virtually no children were predicted to have died
of malaria in this region.
PREDICTING DEATH IN COUNTRIES WITHOUT VITAL REGISTRATION
1047
Table 3 Results of the simultaneously estimated regression models relating proportional mortality outcomes to characteristics of the study designs
and study populations
Variables included in model
Interpretation of effects
P-value
Diarrhoea:pneumonia ratio (‘R2’ = 0.33)
Level of under-5 mortality
As mortality falls from 300 per 1000 to around 165 per 1000, the ratio of diarrhoea
deaths to pneumonia deaths ratio falls by a factor of approximately two. Below this
level, the ratio starts to rise again. At a mortality rate of 26 per 1000, the ratio is
similar to values observed at a rate of 300 per 1000
0.030
Computer algorithm used to determine
definitive cause
This method is associated with a ratio of diarrhoea deaths:pneumonia deaths
approximately two times lower than those obtained using other methods of
determining the definitive cause of death
0.032
Age of the youngest child under mortality
surveillance
Exclusion of the neonatal period from mortality surveillance is associated with a
two- to three-fold increase in the ratio of diarrhoea deaths to pneumonia deaths
0.052
Malaria:pneumonia ratio (‘R2’ = 0.90)
Proportion of the population living in areas
The ratio malaria deaths:pneumonia deaths is approximately four times greater
of endemic, African-pattern falciparum malaria when 100% of the population is exposed to endemic, African-pattern falciparum
malaria than when 50% of the population is exposed. Malaria deaths are negligible
at very low levels of endemicity
0.001
Proportion of births attended by a qualified
professional
Every 10 percentage point increase in the coverage of professionally assisted births is
associated with 19% fewer pneumonia deaths relative to malaria deaths.
0.001
South Asia
This region has a malaria deaths to pneumonia deaths ratio nearly four times higher
than other non-African regions
0.001
Access to safe water
Each 10 percentage point increase in access to safe water is associated with 15%
fewer malaria deaths relative to pneumonia deaths
0.003
Measles:pneumonia ratio (‘R2’ = 0.37)
Measles vaccination coverage rate
Every 10 percentage point increase in the measles vaccination coverage rate is
associated with approximately one third fewer measles deaths relative to
pneumonia deaths
(Mid-)study year
On average, each year associated with a 9% drop in the ratio of measles deaths to
pneumonia deaths
Sub-Saharan Africa
This region associated a ratio of measles deaths to pneumonia deaths approximately
two times greater than other regions
0.001
0.061
0.091
‘Neonatal and other’ cause:pneumonia ratio (‘R2’ = 0.59)
Level of under-5 mortality
As under-5 mortality falls from 300 per 1000 to 100 per 1000 the ratio of ‘neonatal
and other’ cause deaths to pneumonia deaths increases by approximately 60%. This
association becomes much stronger at lower levels of mortality, so that at mortality
levels of 26 per 1000, the ratio is over seven times greater than that observed at a
rate of 100 per 1000
0.001
Computer algorithm used to determine
definitive cause
This method is associated with a ratio of deaths from ‘neonatal and other’ causes to
pneumonia deaths two times lower than when alternative methods are used to
determine the definitive cause of death
0.004
South Asia
This region has a ratio of deaths from ‘neonatal and other’ causes to pneumonia
deaths approximately 35% higher than other regions
0.023
Undetermined:pneumonia ratio (‘R2’ = 0.43)
Method of determining definitive cause
of death
Compared with the use of a standardized diagnostic algorithm, using a single expert,
a panel of experts deciding by majority vote, or an entirely computerized system
is in each case associated with very much greater (44 to 128 times greater) ratio of
deaths from undetermined causes to pneumonia deaths
Level of under-5 mortality
Each 25 per 1000 point fall in the under-5 mortality rate is associated with a 19%
lower ratio of deaths from undetermined causes to pneumonia deaths
Table 4 shows the results of the internal validation. This analysis
indicates that for the group of 38 studies with no missing
covariates, the model successfully reproduced the observed
proportional distribution of deaths by cause. Bias exceeded two
percentage points only for deaths from undetermined and
from ‘neonatal and other’ causes. In relative terms, however,
measles deaths and deaths from undetermined causes were
underestimated by approximately 30%. The substantial
0.001
unexplained heterogeneity of the studies was reflected in large
prediction standard errors, reaching 12 percentage points for
diarrhoea, malaria, and ‘neonatal and other’ causes. If
prediction errors of the same magnitude were associated with
our national (/state) predictions, and assuming that these errors
were independent across countries and across causes of death,
then the 95% CI for the regional distributions would be: subSaharan Africa, diarrhoea 15.5–28.2%, pneumonia 16.1–24.0%,
1048
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
Table 4 Results of the internal validation using jackknife estimation. Based on the 38 studies with no missing covariates
Bias
Observed proportion
of deaths
Predicted proportion
of deaths
Absolute
Relative
Standard deviation
of the residuals
Diarrhoea
22.0%
22.3%
0.3%
1.5%
11.5%
Pneumonia
20.5%
21.4%
0.9%
4.5%
7.0%
8.2%
7.8%
0.4%
5.0%
12.1%
Cause
Malaria
Measles
Neonatal and other
Undetermined
2.7%
1.9%
0.8%
30.7%
3.7%
39.2%
41.5%
2.2%
5.7%
11.5%
7.4%
5.2%
2.3%
30.4%
9.6%
100.0%
100.0%
malaria 17.0–37.0%, measles 1.6–12.8%, ‘neonatal and other’
causes 20.1–33.1%, and undetermined 0.4–6.0%; South Asia,
diarrhoea 16.9–32.8%, pneumonia 17.8–26.4%, malaria
0.1–0.8%, measles 1.0–11.4%, ‘neonatal and other’ causes
40.6–57.3%, and undetermined 1.0–4.7%.
Discussion
This analysis unifies and summarizes best available data about the
distribution of under-5 deaths by cause in countries with
mortality rates of 26 per 1000. Based on population data58 and
reported mortality rates,59 we estimate that these countries
accounted for 86% of all births in the world in 2000, and 98% of
all under-5 deaths. Virtually none of these countries, with the
possible exceptions of Mexico and Suriname, have vital
registration systems complete enough to support valid estimates
of the distribution of causes of death among children under 5.
Current estimates of the distribution of child deaths by cause
must therefore draw on epidemiological studies using verbal
autopsy methods, sometimes supplemented by clinical records.
This study is the first to propose a systematic method for
exploiting the information from these studies that does not
assume that the locations where they were conducted
are representative of entire countries or even supra-national
regions.
Several issues should be kept in mind when interpreting these
results. First, studies of childhood deaths in developing countries
have shown that causes of death established using verbal
autopsy methods are not always consistent with diagnoses based
on more complete clinical data.60–62 In this study, however,
we are interested only in the proportion of deaths attributable to
each cause, and ‘…the fact that there is misclassification, in and
of itself, does not necessarily imply that the resulting verbal
autopsy estimate of the cause-specific mortality fraction will be
inaccurate’ if there are equal numbers of false positives and false
negatives.63 Whether there are counterbalancing numbers of
false positives and false negatives is likely to vary by cause of
death, location, and type of data collection instrument. We did
not adjust our estimates for possible misclassification because we
did not know the technical properties of the instruments used,
and adjusting for misclassification error based on sensitivities
and specificities derived from a validation study population with
a cause of death distribution different from that of the general
population can lead to spurious results.64
Second, the coefficients in our model may be biased. There
may be characteristics of the study populations or study designs
not able to be included in our model that cause uncontrolled
confounding. In this respect, the model can only be an
improvement on previous work limited to bivariate associations
between single-cause mortality and supra-national region65 or
total mortality level.66 The fact that many of the explanatory
variables may have substantial misclassification is of greater
concern, especially those based on sub-national estimates from
surveys conducted in years other than those of the index study.
Misclassification of explanatory variables in complex multivariate regression models can lead to bias of an unpredictable
magnitude or direction.67 Our work benefits, however, from
the particular care that we took to minimize misclassification in
one important predictor variable in our model: the overall level
of under-5 mortality in the different study settings. Our internal
validation using jackknife estimation suggests that the model is
successful in reproducing the observed proportional distribution
of mortality by cause in the intensive mortality studies we
reviewed. External validation using vital registration data is
unfortunately not possible due to the absence of countries in
sub-Saharan Africa or South Asia with an under-5 mortality
level of 26 per 1000 and a complete or near-complete system
of vital registration.
Relatively low proportions of the between-study variance in
the outcome measures could be explained in our model (except
for the malaria:pneumonia ratio, which was well modelled).
Some of the variables that we expected to explain well the
different proportional distributions by cause—such as the
prevalence of undernutrition or maternal literacy—turned out
to be associated with most of the major causes in similar ways,
or to lose statistical significance after adjustment for the overall
under-5 mortality rate. Our findings strongly suggest either that
epidemiological patterns in cause of death are much less regular
than policy makers would hope, or that the details of mortality
study implementation have a major impact on the final results.
We are more certain of the latter: the studies reviewed for this
analysis used many different categorizations of cause of death
and methods for assigning them, and the methods for assigning them were significantly associated with proportional
distributions by cause. We urge investigators to use standardized
PREDICTING DEATH IN COUNTRIES WITHOUT VITAL REGISTRATION
data collection methods such as that developed by WHO/Johns
Hopkins/London School of Hygiene & Tropical Medicine.68 It is
unfortunately the case that until more and better data become
available, our partial understanding of the determinants of local
variation in the proportional distribution of deaths by cause will
limit our ability to make accurate predictions for any given
country. However, if the predictions errors are independent
across countries in the same region, then regional aggregate
statistics will be less severely affected.
Attributing each death to a single cause oversimplifies a reality
in which many children die following multiple illnesses (either
consecutively or concurrently). Some studies6,9–10,14,17,40,44
recorded deaths as due to combinations of single causes
(pneumonia and diarrhoea, for example), but most did not.
More must be learnt about co-morbid events if we are to use
epidemiological profiles to make inferences about the impact of
public health interventions. In some cases, there may have been
a systematic tendency to undercount one illness in the presence
of another. For example, pneumonia may have been
underreported when accompanied by malaria, a tendency likely
to have been accentuated by the diagnostic overlap between
these two diseases.69
The cause of death prediction model presented here was constructed using existing data sets, designed for other objectives.
Only one of the studies included in our analysis30 recorded
deaths from human immunodeficiency virus (HIV)/AIDS, and
our predictions therefore ignore AIDS deaths. However, current
estimates suggest that even in sub-Saharan Africa, HIV/AIDS
causes more than 10% of all child deaths only in 13 severely
affected countries.70 In the future, more data are needed on the
distribution of child deaths by cause in areas with high HIV
prevalence, especially among children and women of childbearing age. Similarly, a lack of standardization forced us to
combine deaths in the neonatal period with all deaths coded as
‘other’. We are therefore unable to predict neonatal deaths
despite their assumed importance, particularly in South Asia.
An important contribution of this analysis is its clear
demonstration of the importance of establishing and supporting
the use of standards for assigning and classifying the causes of
under-5 deaths.
Despite these caveats, the present study provides robust data
on the proportional distribution of under-5 deaths by cause in
sub-Saharan Africa and South Asia. We do not present estimates
for other parts of the world because most of the data on which
our model is based are from these two regions. We find that
malaria, pneumonia, and diarrhoea are still the major killers
in sub-Saharan Africa, and that the latter two causes together
also account for nearly one-half of all child deaths in South
Asia. ‘Neonatal and other causes’ account for over one-half
of child deaths in South Asia, but considerably less in subSaharan Africa. The plausibility of this finding is confirmed
by the fact that four recent Demographic and Health Surveys
(DHS) conducted in South Asia (Bangladesh 1996/97,
Bangladesh 1999/2000, India 1998/99, and Nepall 2001) all
found that deaths in the neonatal period accounted for over
40% of all under-5 deaths, whereas all DHS surveys conducted
in sub-Saharan Africa between 1996 and 2001, with the sole
exception of Mauritania 2000–2001, found that deaths in the
1049
neonatal period accounted for less than 35% of all child deaths (and
often less than one quarter).71 Measles is responsible for relatively
few deaths in both regions, a finding that reflects the fact that in
the 47 studies we reviewed, the median proportion of deaths due
to measles was just 2.2% (interquartile range, 0.0–5.1%). The number of measles deaths predicted by our model is not, however, in
agreement with previous work by Stein et al.72 and Miller.73
It is not the purpose of this paper to compare the resulting
predictions on proportionate child mortality by cause; however, it
is notable that in the mid-1980s, pneumonia and diarrhoea were
considered to be the most important causes of child death in both
sub-Saharan Africa and South Asia.74 Relatively few under-5
deaths in sub-Saharan Africa were attributed to malaria. This
picture has now changed, and malaria is thought to be of similar
magnitude to diarrhoea and pneumonia as a cause of child deaths
in sub-Saharan Africa.75–76 The number and proportion of
deaths due to diarrhoea and pneumonia have decreased, and the
number of deaths due to malaria has remained roughly stable.
Comparisons with other estimation methods are difficult
because the details of previously used methods are not published. Ongoing work of WHO and other groups will utilize
various models, including that described here, to refine regional
mortality estimates. Meanwhile, interventions to control deaths
due to pneumonia, diarrhoea, malaria, and neonatal causes will
be essential if under-five mortality is to be reduced.
Acknowledgements
The authors would like to thank the following investigators,
who kindly supplemented the published data on their studies
with additional information: Dr K Anand, All India Institute
of Medical Sciences; Dr S Arifeen, International Centre for
Diarrhoea Disease Research, Bangladesh; Dr S Awasthi, King
George’s Medical College, Lucknow; Dr A Menezes, Federal
University of Pelotas; Dr U D’Alessandro, Prince Leopold
Institute of Tropical Medicine; Dr J Schellenberg, London
School of Hygiene & Tropical Medicine; Dr C Delacollette, World
Health Organization; Dr J-P Chippaux, Institut de Recherche
pour le Développement; Dr V Fauveau, United Nations
Population Fund; Dr S Hirve, King Edward Memorial Hospital,
Pune; Dr W Huang, Guiyang Medical College; Dr K Kahn,
University of Witwatersrand; Dr F Jalil, Lahore; Dr J Katz, Johns
Hopkins University; Dr H Perry, Hospital Albert Schweitzer. No
specific funding was received by any author or institution for
this work. However, the work was conducted under the
auspices of the Child Health Epidemiology Reference Group,
which operates with the financial support of the Bill and
Melinda Gates Foundation and is coordinated by the
Department of Child and Adolescent Health and Development
of the World Health Organization. Dr J Bryce convened this
group and identified the need for the current analysis to be
undertaken, providing useful feedback at many points. We are
also grateful to Dr B Zaba, London School of Hygiene & Tropical
Medicine, for assistance with demographic modelling, and to
Dr C Boschi-Pinto for useful comments on earlier drafts of this
manuscript. The views represented in this article are those of
the individual authors and do not necessarily represent the
views of their institutions.
1050
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
KEY MESSAGES
•
Small-scale studies of child mortality, with cause of death ascertained by post-mortem interview with the child’s carers,
provide a rich source of data on causes of under-5 death in countries without adequate vital registration systems.
•
Systematic associations can be detected between the proportional distribution of deaths by cause and a number of
characteristics of study populations and designs.
•
The studies suggest that pneumonia, malaria, diarrhoea, and ‘neonatal plus other’ causes are all major causes of death
in sub-Saharan Africa, while diarrhoea, pneumonia, and particularly ‘neonatal plus other’ causes are important in
South Asia.
•
The considerable residual heterogeneity observed in the proportional distributions of under-5 deaths by cause indicates
the need for more and better standardized studies of under-5 mortality in poor countries.
References
1 Lopez AD, Ahmad OB, Guillot M et al. Life Tables for 191 Countries for
2000: Data, Methods, Results. Geneva: World Health Organization, 2001.
(GPE Discussion Paper No. 40).
2 Nannan N, Bradshaw D, Mazur R et al. What is the infant mortality
rate in South Africa? The need for improved data. S Afr Med J 1998;
88:1583–87.
3 Mahapatra P, Chalapatri Rao PV. Cause of death reporting systems in
India: a performance analysis. Natl Med J India 2001;14:129–31.
4 Afari EA, Nkrumah FK, Nakana T et al. Impact of primary health care
on childhood and mortality in rural Ghana: the Gomoa experience.
Cent Afr J Med 1995;41:148–53.
5 Anand K, Kant S, Kumar G et al. Development is not essential to
reduce infant mortality rate in India: experience from the Ballabgarh
project. J Epidemiol Community Heath 2000;54:247–53.
6 Arifeen S, Black RE, Antelmann G et al. Exclusive breastfeeding
reduces acute respiratory infection and diarrhea deaths among infants
in Dhaka slums. Pediatrics 2001;108:E67.
7 Awasthi S, Pande VK, Glick H. Under fives mortality in the urban
slums of Lucknow. Indian J Pediatr 1996;63:363–68.
8 Bang AT, Bang RA, Tale O et al. Reduction in pneumonia mortality
and total childhood mortality by means of community-based
intervention trial in Gadchiroli, India. Lancet 1990;336:201–06.
9 Baqui AH, Black RE, Arifeen SE et al. Causes of childhood deaths in
Bangladesh: results of a nationwide verbal autopsy study. Bull World
Health Organ 1998;76:161–71.
10 Baqui AH, Sabir AA, Begum N et al. Causes of childhood deaths in
Bangladesh: an update. Acta Paediatr 2001;90:682–90.
11 Barreto ICHC, Pontes LK, Corrêa L. Vigilância de óbitos infantis em
~o da autópsia verbal e das
sistemas locais de saúde: avaliaça
informaçóes de agentes de saúde (Portuguese). Rev Panam Salud
Publica 2000;7:303–12.
12 Barros FC, Victora CG, Vaughan JP et al. Infant mortality in southern
Brazil: a population based study of causes of death. Arch Dis Child
1987;65:487–90.
13 Victora CG, Barros FC, Huttly SRA et al. Early childhood mortality in
a Brazilian cohort: the roles of birthweight and socioeconomic status.
Int J Epidemiol 1992;21:911–15.
14 Becker S, Waheeb Y, El-Deeb B et al. Estimating the completeness of
under-5 death registration in Egypt. Demography 1996;33:329–39.
15 Bhatia S. Patterns and causes of neonatal and postneonatal mortality
in rural Bangladesh. Stud Fam Plann 1989;20:136–46.
16 Binka FN, Kubaje A, Adjuik M et al. Impact of permethrin
impregnated bednets on child mortality in Kassena-Nankana district,
Ghana: a randomized controlled trial. Trop Med Int Health
1996;1:147–54.
17 D’Alessandro U, Olaleye BO, McGuire W et al. Mortality and
morbidity from malaria in Gambian children after introduction of an
impregnated bednet programme. Lancet 1995;345:479–83.
18 De Francisco A, Hall AJ, Armstrong-Schellenberg JRM et al. The
pattern of infant and childhood mortality in Upper River Division,
The Gambia. Ann Trop Paediatr 1993;13:345–52.
19 Delacollette C, van der Stuyft P, Molima K et al. Etude de la mortalité
globale et de la mortalité liée au paludisme dans le Kivu montagneux,
Zaïre (French). Rev Epidemiol Sante Publique 1989;37:161–66.
20 Delacollette C, Barutwanayo M. Mortalité et morbidité aux jeunes
âges dans une région à paludisme hyperendémique stable, commune
de Nyanza-Lac, Imbo Sud, Burundi (French). Bull Soc Pathol Exot
1993;86:373–79.
21 Delaunay V, Etard JF, Preziosi MP et al. Decline of infant and child
mortality rates in rural Senegal over a 37-year period (1963–1999).
Int J Epidemiol 2001;30:1286–93.
22 Ekanem EE, Asindi AA, Okoi OU. Community-based surveillance of
paediatric deaths in Cross River State, Nigeria. Trop Geog Med
1994;46:305–08.
23 Fauveau V, Koenig MA, Wojtyniak B. Excess female deaths among
rural Bangladeshi children: an examination of cause-specific mortality
and morbidity. Int J Epidemiol 1991;20:729–35.
24 Ghana VAST Study Team. Vitamin A supplementation in northern
Ghana: effects on clinic attendances, hospital admissions, and child
mortality. Lancet 1993;342:7–12.
25 Greenwood BM, Greenwood AM, Bradley AK et al. Deaths in infancy
and early childhood in a well-vaccinated, rural, West African
population. Ann Trop Paediatr 1987;7:91–99.
26 Hirve S, Ganatra B. Prospective cohort study on the survival
experience of under five children in rural Western India. Indian Pediatr
1997;34:995–1001.
27 Huang W, Yu H, Wang F et al. Infant mortality among various
nationalities in the middle part of Guizhou, China. Soc Sci Med
1997;45:1031–40.
28 Ibrahim MM, Omar HM, Persson, LÅ et al. Child mortality in a
collapsing African society. Bull World Health Organ 1996;74:
547–52.
29 Jaffar S, Leach A, Greenwood AM et al. Changes in the pattern of
infant and childhood mortality in Upper River Division, The Gambia,
from 1989 to 1993. Trop Med Int Health 1997;2:28–37.
30 Kahn K, Tollman SM, Garenne M et al. Who dies from what?
Determining cause of death in South Africa’s rural North-East. Trop
Med Int Health 1999;4:433–41.
PREDICTING DEATH IN COUNTRIES WITHOUT VITAL REGISTRATION
31 Khalique N, Sinha SN, Yunus M et al. Early childhood mortality—a
rural study. J Roy Soc Health 1993;113:247–49.
1051
53 Thompson SG, Sharp SJ. Explaining heterogeneity in meta-analysis: a
comparison of methods. Stat Med 1999;18:2693–708.
32 Khan AJ, Khan JA, Akbar M et al. Acute respiratory infections in
54 Royston P, Altman DG. Regression using fractional polynomials of
children: a case management intervention in Abbottabad District,
Pakistan. Bull World Health Organ 1990;68:577–85.
continuous covariates: parsimonious parametric modelling (with
discussion). Appl Stats 1994;43:429–67.
33 Khan SR, Jalil F, Zaman S et al. Early child health in Lahore, Pakistan:
55 StataCorp. Stata Reference Manual Release 7. Vol. 4, Su–Z. College Station,
X. Mortality. Acta Paediatr 1993;82(Suppl.390):109–17.
Texas: Stata Corporation, 2001, pp. 8–14.
34 Menezes AMB, Victora CG, Barros FC et al. Mortalidade infantile em
56 WHO/UNICEF/The World Bank. State of the World’s Vaccines and
duas coortes de base populacional no Sul do Brasil: tendências e
diferencias (Portuguese). Cad Saude Publica 1996;12(Suppl.1): 79–86.
57 Efron B, Tibshirani RJ. An Introduction to the Bootstrap. Boca Raton:
35 Mølbak K, Aaby P, Ingholt L et al. Persistent and acute diarrhoea as the
Immunization. Geneva, World Health Organization, 2002.
Chapman & Hall/CRC, 1993, pp. 141–51.
leading causes of child mortality in urban Guinea Bissau. Trans Roy Soc
Trop Med Hyg 1992;86:216–20.
58 United Nations Population Division. World Population Prospects: The 2000
36 Mostafa G, Shaikh MAK, van Ginneken JK et al. Demographic
59 UNICEF statistics, child mortality. Available at: http://www.childinfo.
Surveillance System—Matlab. Vol. 28. Registration of Demographic Events—
1996. Dhaka, Bangladesh: International Centre for Diarrhoeal Disease
Research, Bangladesh (ICDDR,B), 1998. (Scientific Report No. 82.)
Revision. New York: United Nations Population Division, 2001.
org/cmr/revis/db2.thm. Accessed 3 May 2003.
60 Kalter HD, Gray RH, Black RE et al. Validation of post-mortem
37 Mtango FDE, Neuvians D, Broome CV et al. Risk factors for deaths in
interviews to ascertain selected causes of death in children. Int J
Epidemiol 1990;19:380–86.
children under 5 years old in Bagamoyo District, Tanzania. Trop Med
Parasitol 1992;43:229–33.
61 Snow RW, Armstrong JR, Forster D et al. Childhood deaths in Africa:
38 Nelson CM, Sutanto A, Gessner BD et al. Age- and cause-specific
childhood mortality in Lombok, Indonesia, as a factor for determining
the appropriateness of introducing Haemophilus influenziae Type b and
pneumococcal vaccines. J Health Popul Nutr 2000;18:131–38.
39 Rahmathullah L, Underwood BA, Thulasiraj RD et al. Reduced
mortality among children in southern India receiving a small weekly
dose of vitamin A. N Engl J Med 1990;323:929–35.
40 Scumacher R, Swedberg E, Diallo MO. Mortality Study in Guinea:
Investigating the Causes of Death for Children Under 5. Arlington, VA:
BASICS II, 2002.
41 Shamebo D, Muhe L, Sandström A et al. The Butajira Rural Health
Project in Ethiopia; mortality pattern of the under fives. J Trop Pediatr
1991;37:254–61.
42 West KP Jr, Pokhrel RP, Katz J et al. Efficacy of vitamin A in reducing
preschool child mortality in Nepal. Lancet 1991;338:67–71.
43 WHO/CHD Immunisation-Linked Vitamin A Supplementation Study
Group. Randomised trial to assess benefits and safety of vitamin A
supplementation linked to immunisation in early infancy. Lancet
1998;352:1257–63.
44 Yassin KM. Indices and sociodemographic determinants of childhood
mortality in rural Upper Egypt. Soc Sci Med 2000;51:185–97.
45 Fishman S, Caulfield LE, de Onis M et al. Underweight status. In:
Ezzati M, Lopez AD, Rodgers A, et al. (eds). Comparative Quantification
of Health Risks: Global and Regional Burden of Disease Attributable to
Selected Risk Factors. Geneva: World Health Organization (in press).
46 INDEPTH Network. Population and Health in Developing Countries. Vol. 1:
Population, Health and Survival at INDEPTH Sites. Ottawa, Canada:
International Development Research Centre, 2002.
47 Brass W, Blacker J. The Estimation of Infant Mortality from Proportions Dying
Among Recent Births. Centre for Population Studies Research Paper 99–1.
London: London School of Hygiene & Tropical Medicine, 1999.
48 David PH, Bisharat L, Hill AG. Measuring Childhood Mortality: A Guide
for Simple Surveys. Amman, Jordan: UNICEF Regional Office for the
Middle East and North Africa, 1990.
49 Salomon JA, Murray CJL. Compositional Models for Mortality by Age, Sex
and Cause. Geneva: World Health Organization, 2001. (GPE Discussion
Paper Series: No. 11.)
50 Katz J, King G. A statistical model for multiparty electoral data. Am Pol
Sci Rev 1999;93:15–32.
51 Aitcheson J. The Statistical Analysis of Compositional Data. New York:
John Wiley & Sons, 1986.
52 Zellner A. An efficient method of estimating seemingly unrelated
regressions and tests for aggregation bias. JASA 1962;57:348–68.
uses and limitations of verbal autopsies. Lancet 1992;340:351–55.
62 Mobley CC, Boerma JT, Titus S et al. Validation study of a verbal
autopsy method for causes of childhood mortality in Namibia. J Trop
Pediatr 1996;42:365–69.
63 Anker M. The effect of misclassification error on reported cause-
specific mortality fractions from verbal autopsy. Int J Epidemiol
1997;26:1090–96.
64 Chandramohan D, Setel P, Quigley M. Effect of misclassification of
causes of death in verbal autopsy: can it be adjusted? Int J Epidemiol
2001;30:509–14.
65 Bern C, Martines J, de Zoysa I et al. The magnitude of the global
problem of diarrhoeal disease: a ten-year update. Bull World Health
Organ 1992;70:705–14.
66 Williams BG, Gouws E, Boschi-Pinto C et al. Estimates of world-wide
distribution of child deaths from acute respiratory infections. Lancet Inf
Dis 2002;2:25–32.
67 Levi M. Errors in the variables bias in the presence of correctly
measured variables. Econometrica 1973;41:985–86.
68 Anker M, Black RE, Coldham C et al. A Standard Verbal Autopsy Method
for Investigating Causes of Death in Infants and Children. Geneva: World
Health Organization, 1999. (WHO/CDS/CSR/ISR/99.4.)
69 World Health Organization. The Overlap in the Clinical Presentation and
Treatment of Malaria and Pneumonia in Children: Report of a Meeting.
Geneva: World Health Organization, 1992 (WHO/ARI/92.23 and
WHO/MAL/92.1065).
70 Walker N, Schwartländer B, Bryce J. Meeting international goals in
child survival and HIV/AIDS. Lancet 2002;360:284–89.
71 Measure DHS+. Demographic and health surveys, publications & press.
Available at: http://www.measuredhs.com/pubs/. Accessed 3 May 2003.
72 Stein CE, Birmingham M, Kurian M, Duclos P, Strebel P. The global
burden of measles in the year 2000—a model that uses countryspecific indicators. J Infect Dis 2003;187(Suppl.1):S8–S14.
73 Miller MA. Introducing a novel model to estimate national and global
measles disease burden. Int J Infect Dis 2000;4:14–20.
74 Lopez AD. Causes of death in industrial and developing countries:
estimates for 1985–1990. In: Jamison DT, Mosley WH, Measham AR,
Bobadilla JL (eds). Disease Control Priorities in Developing Countries.
New York: Oxford University Press, 1993, pp. 35–50.
75 Snow
RW, Craig M, Deichmann U, March K. Estimating
mortality, morbidity and disability due to malaria among Africa’s nonpregnant population. Bull World Health Organ 1999;77:624–40.
76 Mathers CD, Stein C, Fat DM et al. Global burden of disease 2000.
Version 2 methods and results. WHO GBD documentation (http://
www.who.int/evidence, accessed 14 March 2003).