IJE vol.33 no.6 © International Epidemiological Association 2004; all rights reserved.
Advance Access publication 19 August 2004
International Journal of Epidemiology 2004;33:1293–1301
doi:10.1093/ije/dyh182
Monitoring trends in under-5 mortality rates
through national birth history surveys
EL Korenromp,1 F Arnold,2 BG Williams,3 BL Nahlen4 and RW Snow5
Accepted
23 February 2004
Background We assessed whether Demographic and Health Surveys (DHS), a large and
high-quality source of under-5 mortality estimates in developing countries,
would be able to detect reductions in under-5 mortality as established in global
child health goals.
Methods
and
Results
Mortality estimates from 41 DHS conducted in African countries between 1986
and 2002, for the interval of 0–4 years preceding each survey (with a mean time
lag of 2.5 years), were reviewed. The median relative error on national mortality
rates was 4.4%. In multivariate regression, the relative error decreased with
increasing sample size, increasing fertility rates, and increasing mortality rates.
The error increased with the magnitude of the survey design effect, which
resulted from cluster sampling. With levels of precision observed in previous
surveys, reductions in all-cause under-5 mortality rates between two subsequent
surveys of 15% or more would be detectable. The detection of smaller mortality
reductions would require increases in sample size, from a current median of 7060
to over 20 000 women. Across the actual surveys conducted between 1986 and
2002, varying mortality trends were apparent at a national scale, but only around
half of these were statistically significant.
Conclusions The interpretation of changes in under-5 mortality rates between subsequent
surveys needs to take into account statistical significance. DHS birth history
surveys with their present sampling design would be able to statistically confirm
under-5 mortality reductions in African countries if true reductions were 15% or
larger, and are highly relevant to tracking progress towards existing international
child health targets.
Keywords
Child mortality/*trends, infant mortality/*trends, time factors; Africa/
epidemiology; programme monitoring, demographic and health surveys,
developing countries/*statistics & numerical data
All-cause under-5 mortality is a key health outcome in
developing countries, and the reversal in Africa during the 1990s
of the mortality decline apparent since the 1960s is the subject of
1 World Health Organization, Roll Back Malaria Dept., Avenue Appia 20,
CH 1211—Geneva 27, Switzerland.
2 ORC Macro, 11785 Beltsville Drive, Calverton, MD 20705 USA.
3 World Health Organization, StopTB Department, Avenue Appia 20,
CH 1211—Geneva 27, Switzerland.
4 World Health Organization, Roll Back Malaria Department, Avenue Appia
20, CH 1211—Geneva 27, Switzerland.
5 KEMRI Wellcome Trust Collaborative Programme, 00100 GPO, P.O. Box
43640, Nairobi, Kenya and Centre for Tropical Medicine, University of
Oxford, John Radcliffe Hospital, Oxford, OX3 9DU, UK.
Correspondence: Dr EL Korenromp, World Health Organization, Roll Back
Malaria Department, Avenue Appia 20, CH 1211—Geneva 27, Switzerland.
E-mail: [email protected].
much concern.1–7 The reversal has been attributed mainly to the
human immunodeficiency virus (HIV) epidemic, although HIVrelated mortality alone cannot fully explain this trend.6,7
Various major health programmes and initiatives focus on
under-5 mortality. Most UN member states have agreed to the
UN Millennium Development Goal (MDG) of reducing the
under-5 mortality rate by two-thirds between 1990 and 2015.8
Under-5 mortality is included among the indicators proposed by
the Global Fund against AIDS, TB, and Malaria.9 With malaria
causing, or contributing to, over 20% of deaths in African
children,10–12 the all-cause under-5 mortality rate is also
increasingly regarded as an important indicator of the impact of
malaria control.13
In Africa, where civil registration is in most countries
notoriously inadequate,14 a main source of statistics on all-cause
under-5 mortality is the Demographic and Health Surveys
1293
1294
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
(DHS), which estimate infant and childhood mortality rates
through nation-wide sample surveys of women aged 15–49
years.15 DHS were initiated in the mid-1980s in order to
monitor key population, health and nutrition outcomes,
notably on reproductive health, in populations where such
information was lacking from routine health systems and from
vital registration. The surveys have subsequently been
expanded to cover a wide range of topics related to population
and health programmes. They are carried out approximately
every 5 years in an increasing number of developing countries.
We assessed what magnitude of change in under-5 mortality
rates in African countries would be statistically detectable
through DHS surveys, and what sample size would be required
to detect expected or targeted levels of mortality reductions. The
results are discussed in the light of the urgent need to provide
robust measures of public health impacts in accordance with
new international targets and increased investment in disease
control.
Demographic and Health Surveys
DHS obtain nationally representative outcomes by using a
two-stage cluster design.15 In the first stage, a subset of
geographical clusters (usually census enumeration areas) is
selected proportional to population size. In the second stage, a
sample of households is selected from a complete household
listing in the selected clusters. In some cases, the selection of
clusters is a three-stage process. Cluster sampling, as opposed to
taking a random, completely dispersed sample of all households
in the country, is logistically necessary to perform the surveys
within a reasonable time period at an acceptable cost.
Sampling usually occurs in proportion to the size of the
population living in each cluster, with an equal number of
households being chosen in each cluster within a sampling
domain, so that the sample design is self-weighting. However,
certain clusters are sometimes oversampled, in order to allow a
sufficiently large sample to yield reliable subnational estimates.
To produce outcomes for the urban and rural parts of countries
separately, urban areas may be oversampled in countries where
these areas constitute only a small part of the total population,
e.g. in Cameroon in 1991 and 1998 and in Uganda in 1988. In
some surveys, e.g. Malawi 2000 and Uganda 2000/2001, certain
areas were oversampled because of programmatic demands.
Typically, the sample size is between 15 and 30 households for
urban clusters, and between 30 and 40 households for rural
clusters.
Survey outcomes are affected by two types of sampling errors:
(1) the statistical error due to the limited sample size, and (2)
the design effect, which represents the factor by which the
cluster-based sampling compounds this error. A design effect of
1.2 means that the total error is 1.2 times higher than it would
have been if a simple random sample, without clustering, had
been chosen. The design effect depends on the number
of households per cluster and on the extent to which the
outcome of interest (e.g. child mortality) varies within and
between clusters. Therefore, for a given survey design, the
design effect differs between outcomes and between countries.
DHS surveys provide details of the survey design as well as the
design effects for a list of key outcomes in appendices of their
final reports.
Mortality data from DHS
DHS estimate under-5 mortality rates from complete birth
histories, based on reports from mothers on the survival of their
children. The direct estimation technique is based on a life table
approach: probabilities of dying are computed from reported
dates of birth and death and the numbers of children of a
particular age exposed to the risk of dying during a specified
period.16,17 Each survey produces national mortality rates for
the intervals 0–4, 5–9, and 10–14 years preceding the survey.
These estimates can be considered to indicate the mortality rates
at a mean of 2.5 years, 7.5 years, and 12.5 years preceding the
survey, respectively. The birth history data for the less recent
periods are thought to be of lower quality because event
omission and misreporting of date of birth and age at death for
deceased children are likely to occur more frequently at longer
durations of recall. This leads to progressive underestimation of
mortality rates with increased time preceding the survey, as has
been demonstrated by comparing rates for a given calendar
period between recent and less recent surveys.17 For this
reason, our analyses concentrate on mortality rates for the
5 years preceding the survey.
For the purpose of country-level mortality monitoring
we focus on national-level outcomes. Figure 1 shows African
countries where national estimates of under-5 mortality rates
are available from one, two, or more surveys. Table 1 lists the
standard errors on mortality, for selected surveys in subSaharan Africa. The median error is 6.45; at a median mortality
rate of 151 per 1000 births, this translates to a relative error
of 4.4% (range: 2.5%, 7.6%). The median design effect for
mortality is 1.30 (range: 1.08, 2.07), at an overall median
number of households per cluster (pooled over urban and rural
strata) of 25.7. These surveys sampled a median of 7060 women
(range 3200 to 15 367); the median total fertility rate (estimated
cumulative lifetime births per woman at current fertility rates)
across the surveys was 5.8 (range: 4.0, 7.4).
Figure 1 Number of Demographic and Health Surveys (DHS) surveys
conducted in African countries up to 200215
MONITORING OF CHILD MORTALITY
1295
Table 1 Under-5 mortality rates and statistical errors estimated from Demographic and Health Surveys (DHS) conducted in selected
sub Saharan African countries between 1986 and 2002
Survey country, year
Benin 1996
a
5q0
Women
interviewed
Mean
households
per cluster
Design
effectb
Total
fertility
ratec
SEd on
5q0
Relative
SE on 5q0
166.5
5491
23.8
1.15
6.0
6.44
3.9%
Benin 2001
159.8
6216
25.0
1.15
5.6
5.94
3.7%
Cameroon 1998
150.7
5501
24.7
1.41
4.8
8.10
5.4%
Central African Republic 1994
157.5
5884
26.7
1.18
5.1
6.45
4.1%
Chad 1996/7
194.2
7454
29.4
1.31
6.4
6.66
3.4%
Cote d’Ivoire 1994
149.5
8099
25.8
1.38
5.2
6.12
4.1%
Eritrea 1995
136.2
5054
30.1
1.31
6.1
7.50
5.5%
Ethiopia 2000
166.1
15 367
27.2
1.53
5.5
5.49
3.3%
Gabon 2000
88.6
6183
27.2
1.19
4.2
5.94
6.7%
Ghana 1988
154.7
4488
21.5
1.33
6.4
7.93
5.1%
Ghana 1993
119.4
4562
15.4
1.09
5.2
6.18
5.2%
Ghana 1998
107.6
4843
15.9
1.08
4.4
6.42
6.0%
Kenya 1989
89.8
7150
5.9
1.62
6.7
6.30
7.0%
Kenya 1993
96.1
7540
16.9
1.29
5.4
5.24
5.5%
Kenya 1998
111.5
7881
17.9
1.47
4.7
7.17
6.4%
Madagascar 1997
159.2
7060
29.4
1.60
6.0
7.77
4.9%
Malawi 1992
233.8
4849
25.8
1.21
6.7
8.54
3.7%
Malawi 2000
188.6
13 220
27.6
1.20
6.3
4.70
2.5%
Mali 1987
247.0
3200
15.7
1.72
6.7
13.5
5.5%
Mali 1995/96
237.6
9704
31.7
1.36
6.7
5.90
2.5%
Mali 2001
229.1
12 849
34.1
1.64
6.8
6.47
2.8%
Mozambique 1997
200.9
8779
28.1
2.07
5.2
9.79
4.9%
Nigeria 1990
192.8
8777
33.4
1.97
6.0
10.64
5.5%
Nigeria 1999
140.2
8201
19.8
1.31
4.7
6.65
4.7%
Rwanda 2000
196.2
10 421
23.0
1.41
6.2
6.67
3.4%
Senegal 1986
194.7
4415
9.8
1.20
6.6
7.93
4.1%
Senegal 1992/93
131.8
6310
14.4
1.25
6.2
5.96
4.5%
Senegal 1997
139.1
8593
15.7
1.24
6.0
5.61
4.0%
Sudan 1990
123.7
5860
16.5
1.17
4.6
5.05
4.1%
Tanzania 1992
140.9
9238
26.6
1.56
6.2
6.55
4.7%
Tanzania 1996
136.5
8120
24.9
1.30
5.8
5.84
4.3%
Tanzania 1999
146.6
4029
21.7
1.36
5.6
9.08
6.2%
Togo 1998
146.3
8569
28.0
1.19
5.2
5.42
3.7%
Uganda 1988
177.0
4730
17.1
1.20
7.4
6.95
3.9%
Uganda 1995
147.4
7070
27.4
1.36
6.9
6.09
4.1%
Uganda 2000/01
151.5
7246
26.5
1.46
6.9
6.66
4.4%
Zambia 1992
191.2
7060
25.7
1.23
6.5
6.79
3.6%
Zambia 1996
196.6
8021
25.7
1.08
6.1
5.47
2.8%
70.6
4201
17.7
1.13
5.3
5.39
7.6%
Zimbabwe 1988
Zimbabwe 1994
77.1
6128
28.1
1.13
4.3
5.02
6.5%
Zimbabwe 1999
102.1
5907
30.3
1.20
4.0
6.76
6.6%
Median
150.7
7060
25.7
1.30
5.8
6.45
4.4%
a q = The probability of dying before the fifth birthday, expressed as deaths per 1000 births.
5 0
b The design effect represents the factor by which the cluster sampling contributes to the total SE on the under-5 mortality rate (see text).
c The cumulative number of births per woman aged 15–49 years (estimated from a hypothetical rectangular distribution of the age-specific fertility rates over
the 3-year period preceding the survey, to represent the number of births a woman would have had in her lifetime if at each age she had the average number
of births that women in the survey had over the preceding 3 years).
d Standard error.
Not included are surveys for which error data were not available.
Source: DHS final reports, appendices, and unpublished data (ORC Macro, MEASURE DHS+).
1296
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
Mortality trends in actual DHS
Figure 2 shows time trends in under-5 mortality rates for
African countries in which three or more DHS provided
mortality statistics between 1986 and 2002. The existence of
differences in mortality among the available surveys for each
country was evaluated with a 2 test, using the standard errors
published in the final reports. This test detects the presence of
any deviation from a constant mortality rate, whether this was
an increase or decrease between any two of the three surveys,
an increase or decrease continuing over both intervals, an
increase followed by a decrease or vice versa. If the 2 test
revealed any difference, its magnitude and direction between
any two surveys was subsequently evaluated by two-tailed
t-tests (assuming the errors on mortality rates to be normally
distributed). For this analysis, Malawi and Nigeria, each with
two surveys, were also included. Malawi was of special interest
as the single African country that adopted early in the 1990s
a changed, effective antimalarial drug policy, after which it
showed a large decline in under-5 mortality. Nigeria was
included for being the most populous African country.
All countries except Kenya, Mali, and Tanzania revealed
significant variation in mortality rates (Table 2). Mortality
increased over the three surveys in Zimbabwe (P 0.001). In
Ghana, mortality decreased significantly over the first interval
(P 0.001), and non-significantly over the second interval. In
Senegal and Uganda, mortality fell over the first interval (P 0.001 and P = 0.0014) and increased thereafter, but the latter
trends were not significant. In Zambia, in contrast, mortality
fell exclusively over the second interval. In both countries with
only two surveys, Malawi and Nigeria, mortality fell
significantly (P 0.001 in both cases).
Scenarios of mortality reductions
Figure 2 Under-5 mortality (deaths before age five per 1000 births, or
5q0) by calendar year for sub-Saharan African countries with three
or more Demographic and Health Surveys (DHS) surveys available.
Error bars represent standard errors. Statistics are for the 0–4 years
preceding each survey
DHS calculate the errors on 5q0 estimates from the survey
databases. Rather than calculating the error from its
determinants such as the number of clusters (as can be done for
outcome measures less complex than mortality), the error is
estimated using the Jackknife repeated replication method.18 In
the Jackknife method, mortality rates are repeatedly calculated
for replications of the dataset that each include all but one
(alternating) cluster; the standard error on mortality rates is
then calculated using a simple formula (described in the
appendices of the final reports). This method is needed because
the error depends not only on the numbers of sampled clusters,
enumeration areas from which the clusters were sampled,
Table 2 Time trends in under-5 mortality (5q0) in sub-Saharan African countries with three or more Demographic and Health Surveys (DHS)
surveys conducted between 1986 and 2002, and Malawi and Nigeria
5q0 levels (95% CI)
P-value for overall
differences
(irrespective of
pattern)a
First survey
Second survey
Third survey
Ghana
155 (139, 170)
119 (107, 132)
108 (95, 120)
0.001*
Kenya
90 (77, 102)
96 (86, 106)
112 (97, 126)
0.069
Malawi
234 (217, 250)
188 (179, 198)
NA
NA
Mali
247 (221, 273)
238 (226, 249)
229 (216, 242)
0.40
P-value for change between
subsequent surveysb
1st interval
2nd interval
0.001*
0.18
Not evaluated
0.001*
NA
Not evaluated
Nigeria
193 (172, 214)
140 (127, 153)
NA
NA
0.001*
NA
Senegal
195 (179, 210)
132 (120, 143)
139 (128, 150)
0.001*
0.001*
0.37
Tanzania
141 (128, 154)
137 (125, 148)
147 (129, 164)
0.63
Uganda
177 (163, 190)
147 (135, 159)
151 (138, 164)
0.003*
0.0014*
0.65
Zambia
191 (178, 205)
197 (186, 207)
168 (156, 180)
0.0018
0.54
0.0006
71 (60, 81)
77 (67, 87)
102 (89, 115)
0.001*
0.38
0.003*
Zimbabwe
Not evaluated
Statistics for 0–4 years preceding survey. * = significant at P = 0.025 level; NA = not available.
a Evaluated by 2-test.
b Evaluated by two-tailed t-test, assuming the errors on estimated mortality rates to be normally distributed. Only evaluated if overall difference (5th column)
was significant at P = 0.05.
MONITORING OF CHILD MORTALITY
births reported on, and deaths, but also on the extent of intraand inter-cluster variation that occurred in mortality rates in
the specific survey population.
To predict the extent of statistical error under hypothetical
scenarios of changing mortality and what sample size would be
required to detect such trends, we described the errors in all
African surveys available as of September 2003 as a function of
the baseline mortality rate, design effect, and the number of
women interviewed, as published in final reports of these
surveys. In addition, the total fertility rate was included as an
approximation of the number of births that interviewed women
had in the preceding 5 years, and for which under-5 mortality
was evaluated. A weighted least squares multivariate model was
fitted in SPSS version 10.0.7 (SPSS Inc., 1989–1999). Each
survey was given an equal weight in the multivariate analyses.
The four determinants were included as continuous variables; a
logistic transformation was applied on 5q0 and the effects of the
survey sample size and of the total fertility rate were linearized
by using the reciprocal of their square roots.
The model obtained a good fit: across the 41 surveys, the four
transformed variables together explained 94% of the variation
in statistical error (Appendix). The error increased slightly with
the baseline mortality rate (P 0.001, Figure 3a). This increase
was less than proportional, so that the relative error, in contrast,
1297
decreased with the baseline mortality rate (not shown). The
error decreased with the number of women interviewed (P 0.001, Figure 3b) and with the total fertility rate (P = 0.001,
Figure 3c). Although both these variables are associated with
the number of births for which survival is evaluated, the
influence of fertility level was less marked because the surveys
differed less in fertility levels than in sample size. Large design
effects increased the error (P 0.001, Figure 3d). The effects of
all variables were independent from one another.
If the standard error were a linear function of the included
parameters, the proportion of variation explained would be
expected to be 100%. The reason it is slightly less (94%) may
be that parameter effects are in reality not linear, or that some
relevant aspect of design that varied between surveys was not
fully captured in the design effect. In addition, the total fertility
rate does not exactly reflect the number of births per woman for
which mortality was evaluated.
The model allowed us to predict the standard error for
hypothetical surveys of given sample size and baseline mortality
rate. For this evaluation we assumed a fixed design effect of
1.30, the median across all actual surveys (Table 1). The
statistical significance of mortality trends over subsequent
hypothetical surveys with a predicted standard error were
evaluated by assuming mortality rates to be normally
Figure 3 Standard errors (SE) on the under-5 mortality estimates (5q0) for the interval of 0–4 years preceding each Demographic and Health
Survey (DHS). Squares are SE calculated for individual surveys conducted in sub-Saharan African countries up to 2002; lines are covariateadjusted predictions from a multivariate model (see Appendix). (a) Error as a function of the baseline mortality level, for a median sample size of
7060 women, design effect of 1.30, and total fertility rate of 5.8; (b) error as a function of the number of women interviewed, for a median
under-5 mortality level of 151 per 1000 births, design effect of 1.30, and total fertility rate of 5.8; (c) error as a function of the total fertility rate,
for a median sample size of 7060 women, mortality level of 151 per 1000 under-5s and design effect of 1.30; (d) error as a function of the design
effect, for a median sample size of 7060 women, mortality level of 151 per 1000 under-5s, and total fertility rate of 5.8
1298
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
Figure 4 Survey sample size (number of women interviewed)
required to statistically detect certain reductions in under-5 mortality,
as a function of the baseline under-5 mortality rate (in deaths
per 1000 births). Significance evaluated two-tailed at P = 0.025 level,
assuming normal distributions. Estimated based on multivariate model
presented in Figure 3 and Appendix, assuming a fixed design effect of
1.30 and a total fertility rate of 5.8 children per woman. The lower left
corner represents combinations of mortality level and reduction that
would require a sample size of 16 000 women. Mortality reductions
smaller than 8% could not be evaluated, because the required sample
size would exceed the upper limit from actual Demographic and Health
Survey (DHS) that were used for error predictions
sizes may be insufficient (Figure 4). Precision decreases with
decreasing baseline mortality rates, so that if under-5 mortality
in sub-Saharan Africa were indeed to decline from its current
level, increasingly larger sample sizes would be required to
monitor subsequent trends.
The international community has established new child
survival targets over the next two decades and signed
commitments to achieve these goals.8 The required two-thirds
reduction in the under-5 mortality rate by 2015 translates to
a reduction by 40% every 5 years from 2004 onwards.
Encouragingly, all national DHS surveys are sufficiently
powerful to detect this change. For Africa, a considerable part
of the expected mortality reduction is to come about by
improved malaria control, under the Roll Back Malaria
strategy20 and continued financial support from the Global
Fund against AIDS, TB, and Malaria.21 Based on evidence from
African malaria intervention trials,22–24 a 15% reduction in allcause under-5 mortality is a realistic (minimum) expectation for
the impact of improved malaria control in countries with
increased distribution of insecticide-treated mosquito nets and
increased coverage of malaria cases with prompt treatment with
effective antimalarial drugs. Our analysis thus supports the
proposal that the all-cause under-5 mortality rate as measured
by national surveys should be a key indicator of the
epidemiological impact of Roll Back Malaria.13
Limitations
distributed. Detection of a mortality rate reduction of 20%
would require a sample size of 6700 in case of a baseline
mortality rate of 70 per 1000 and a sample size of 1450 women
at a baseline mortality rate of 250 per 1000. Not surprisingly,
larger sample sizes would be required with lower baseline
mortality rates, which is graphically illustrated in Figure 4.
Figure 4 also shows that most of the recently completed DHS in
sub-Saharan Africa, with sample sizes of between 4000 and
8000 (Table 1) would be able to pick up a 15–20% mortality
reduction. To statistically detect a mortality rate reduction of
only 10% would, however, require a sample size of between
6500 and 14 800 (Figure 4), approaching the maximum among
actual recent surveys (Table 1).
Discussion
Our review of direct mortality rates from DHS conducted
between 1986 and 2002 in sub-Saharan African countries
confirmed the existence of mortality trends in many countries
(Figure 2). However, not all trends between subsequent surveys
were statistically significant (Table 2), a fact that is frequently
ignored in trend analyses.3–5,19 Although exact errors on
survey-based mortality rate estimates require a sophisticated
calculation on the raw databases, the error can be estimated
based on the published baseline mortality rate, design effect,
numbers of women sampled and the total fertility rate as an
approximation of the births per woman (Figure 3). Using such
error estimates, it appears that the DHS approach in its current
sampling design and with under-5 mortality levels described
during the 1990s in sub-Saharan Africa would be able to
statistically confirm mortality reductions at the national level of
15% or more; to detect more subtle reductions, current sample
A number of limitations must be recognized to the presented
evaluations of mortality time trends and the statistical power of
surveys. Survey-based estimates of under-5 mortality for the
period 0–4 years preceding the survey refer to a midpoint of
2.5 years before the survey. The detection of the mortality
impact of a health programme will therefore be delayed by an
average of 2.5 years; in other words, only 2.5 years after an
effective programme had been implemented could a mortality
impact be reliably attributed to expanding coverage of
interventions. Similarly, survey data from a current year could
serve as the baseline measurement for the evaluation of a
programme that started 2.5 years ago.
At subnational levels, the detection of under-5 mortality time
trends would be less powerful than we have described for the
national level (Figure 4). To compensate for the smaller sample
size, DHS typically provide mortality outcomes for regions,
provinces, and urban or rural areas for a 10-year interval
preceding the survey rather than the 5-year interval used for
national estimates. These reports at sub-national levels are often
used to highlight spatial inequities, but again often without
any consideration of the statistical significance of eventual
differences. Even so, the relative errors on subnational rate
estimates are large: a median of 7.1% of the mortality rate for
urban areas, 3.9% for rural areas, and 7–8% for provinces and
regions (not shown). To statistically detect changes between
surveys for subnational areas would therefore require sample
sizes of at least 35% larger than those shown in Figure 4. More
robust techniques or sampling is required to provide reliable
estimation of subnational mortality trends. These might include
demographic surveillance systems (DSS) which monitor
mortality prospectively over time in defined populations,11,25
although these systems often suffer from a lack of representativeness of the populations covered.25
MONITORING OF CHILD MORTALITY
Similarly, the errors of mortality rates for specific age groups
among the under-5s are larger that those presented here for all
under-5s pooled. For example, for the most recent 0–4 year
interval, the median relative standard error was 5.6% for infant
mortality rates across surveys reported in Table 1 and 6.6% for
child (1–4 years) mortality rates. The magnitude of these
sampling errors dictates that 13–22% larger sample sizes would
be required than for total under-5 mortality rates.
Our analysis did not address non-sampling errors in the birth
history surveys. It is nevertheless likely that non-sampling
errors on under-5 mortality estimates are substantial in some
surveys17,26,27—and these cannot be reduced by increasing
sample size. Although interviewers in DHS are extensively
trained to probe for all births and deaths, and to elicit accurate
information about dates of birth and ages at death, omission of
deaths and misreporting of ages and dates of birth are causes of
concern. Mothers may not report all of their births, particularly
for children born long before the survey and who died at a very
young age. Interviewers may fail to record all deaths in order
to avoid asking potentially uncomfortable questions about
the pregnancy, delivery, and feeding practices for children who
have died. Age inflation may be a problem in some surveys,
particularly in countries where mothers do not know the exact
age of their children. In those countries, some interviewers
might increase the ages of young children to avoid a long set of
questions on maternal and child health, as well as height and
weight measurements, that are restricted to young children.
However, the impact of age displacement on estimates of
childhood mortality has been found to be small or negligible.17
Overall, non-sampling errors may substantially distort mortality
rate estimates in some surveys, although if these errors remain
similar across surveys, the effect on time trends will be small. A
potential bias that will affect time trends is the missing of
children whose mothers have died because birth history data
are limited to the biological children of living women. Since
mortality may be higher among orphaned children,28–31 birth
histories may underestimate overall mortality. This is of
particular concern in countries in Southern and Eastern Africa
with high or rising HIV-related mortality among young women.
1299
mortality rates would add to the error in survey-based all-cause
mortality estimates, decreasing the statistical power to evaluate
trends in non-HIV mortality.
The DHS data represent one of the sources for the under-5
mortality estimation by UNICEF that is agreed upon as one of the
benchmarks of the Millennium Development Goals.33 The latter
estimation takes into account, besides DHS, estimates of mortality
rates from censuses, vital registration, and non-birth history
surveys, such as the Multiple Indicator Cluster Surveys.34 This
data synthesis assigns different weights to the different data and
estimation methods based on their respective quality and validity,
as judged by the analysts. Error estimates are typically not
available for the indirect mortality estimates, but sampling errors
in the non-birth history surveys are likely to be at least as large as
those in DHS surveys, considering several additional biases
inherent to the indirect method. Non-sampling errors can be
substantial in all data sources, and they might introduce spurious
time trends if certain sources with specific biases contribute
proportionally more in certain time periods. Although the
syntheses of all available data might reduce the sampling error in
the resulting overall time trend, it is our belief therefore that the
overall error in such combined estimates is on the same order of
magnitude as we presented here for DHS specifically.
Implications for future surveys
An obvious approach to improving the statistical precision to
detect under-5 mortality transitions would be to increase
sample size. This is apparent among more recent DHS surveys
in sub-Saharan Africa: 8 DHS conducted in 2000 or later had an
average sample size of 9904 women, compared with 6225 over
the 51 surveys conducted before 2000 (Figure 5). As is implicit
in our scenario analysis (Figure 4) from the use of a fixed design
effect, increasing the sample size largely takes the form of more
clusters, rather than more households per cluster. Increasing
the number of clusters sampled is statistically and logistically
preferable to increasing the number of households per cluster,
in which case the gain in statistical power would be less than
shown in Figure 4, owing to an increased design effect.
Explanation and interpretation of trends
The interpretation of statistically detectable mortality changes in
terms of specific diseases or health interventions was beyond the
scope of this analysis. The most universal current determinant of
changing child mortality in Africa is the increase in HIV-related
deaths;7 in evaluations of the impact of major child health
programmes such as the Integrated Management of Childhood
Illnesses (IMCI),32 Roll Back Malaria, or diarrhoeal management
programmes, increasing HIV mortality would confound (deflate)
the apparent impact. The proportion of under-5 mortality
attributable to HIV can be estimated from HIV prevalences in
antenatal clinic attenders and subtracted from all-cause mortality.
For the year 1999, this estimation has been undertaken for
all countries in sub-Saharan Africa.7 HIV infection caused an
estimated 7.7% of under-5 deaths, as compared with 2% in 1990.
Where HIV prevalence estimates are available for other years, the
changing burden of HIV-related under-5 mortality could similarly
be estimated and subtracted from corresponding child mortality
surveys to obtain the trend in non-HIV mortality. However,
uncertainties inherent to the estimation of HIV-related under-5
Figure 5 Sample size (number of women interviewed) of DHS
conducted in sub-Saharan African countries up to 2002, as a function
of calendar year. Diamonds: actual surveys, curve: overall trend (R2 =
0.24, P 0.001)
1300
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
Large-scale surveys such as the DHS are expensive
undertakings and a balance always has to be made between
statistical considerations and financial and personnel costs as
well as operational constraints. It is of note that mortality, being
a probability event that fortunately, even in Africa, affects only
a minority of children, requires a bigger sample size than survey
indicators that measure a ‘prevalence’ among all children,
such as the coverage of all children with mosquito nets,35 or the
prevalence of anaemia. It might therefore be conceivable
to increase sample size exclusively for the most demanding
submodules of the total survey, e.g. with a scaled-down
questionnaire that would be administered to additional clusters
and focused only on child mortality (in a shortened birth
history) and other rare events.
An alternative suggestion has been to increase the frequency
of surveys (e.g. every 3 years) while retaining the current
number of clusters. More frequent surveys would help the
monitoring of outcomes that can change rapidly over time and
that are measurable very precisely, such as the coverage with
mosquito nets among under-5s in response to a malaria control
programme.35 For mortality rates, however, increasing survey
frequency is probably not the most efficient way to increase the
statistical power. Mortality reductions that occurred between
pairs of subsequent surveys would be smaller, because the
group of under-5s contributing to subsequent 5-year intervals
would start overlapping. As a consequence, an even larger
sample size would be required (see Figure 4, in which it is of
note that the sample size requirements hold true irrespective of
the interval between the surveys).
With increasing investment in child survival initiatives in
Africa there is an increasing need to ensure that targets set by
the international community are monitored. Statistical
precision, largely dependent upon sample size, has often been
ignored in the presentation of under-5 mortality trends. Birth
history surveys will continue to be the benchmark to monitor
under-5 mortality rates. In their current design, DHS surveys in
Africa should provide enough power to detect mortality rate
reductions at national levels of 15% or larger within time scales
defined by international partners, and they are thus highly
relevant to tracking progress towards existing international
child health targets. The changing dynamic of mortality needs
to be constantly reassessed in terms of the ability of these
surveys to detect temporal changes: increasing HIV prevalence
will blunt impacts achieved through other disease-specific
interventions and substantial declines in mortality rates would
merit increased cluster sample size over time.
Acknowledgements
We thank Jerry Sullivan, Rick Steketee, Allan Hill, Alfredo
Aliaga, Ruilin Ren, and the Roll Back Malaria Monitoring and
Evaluation Reference Group (MERG) Task Force on mortality
for comments on the manuscript. We are grateful to Ladys Ortiz
for providing tabulations of DHS statistics, and to Eleanor
Gouws for statistical advice. This work received financial
support from the Wellcome Trust, UK, the Kenyan Medical
Research Institute, the Bill & Melinda Gates Foundation, and
the MEASURE project which receives support from the United
States Agency for International Development (USAID) under
contract numbers HRN-A-00–97–00018–00 and HRN-A00–97–00019–00. Bob Snow is supported by The Wellcome
Trust (UK) as part of their Senior Fellows program (#058992).
KEY MESSAGES
•
Demographic and Health Surveys (DHS), conducted at 5-year intervals in an increasing number of countries, are
a main source of high quality estimates of under-5 mortality rates in African countries.
•
In interpreting under-5 mortality rates and trends in mortality rates from DHS, statistical precision and
significance must be taken into account. Across actual surveys conducted between 1986 and 2002, only half of
the trends observed in mortality rates at national levels were statistically significant.
•
With levels of precision observed in previous African DHS, reductions in all-cause under-5 mortality rates
between successive surveys of 15% or more are detectable. This is highly relevant to tracking progress towards
international child health targets; the detection of smaller mortality reductions would, however, require increases
in sample size, from a current median of 7060 to over 20 000 women.
References
1 Foster SO. Monitoring child survival programmes in Africa: the Africa
Child Survival Initiative. Background. Int J Epidemiol 1993;22
(Suppl.1):S2–7.
2 Ewbank DC. A strategy for evaluating the mortality impact of child
4 Rutstein SO. Factors associated with trends in infant and child
mortality in developing countries during the 1990s. Bull World Health
Organ 2000;78:1256–70.
5 Timaeus IM. Impact of the HIV epidemic on mortality in sub-Saharan
Africa: evidence from national surveys and censuses. AIDS
1998;(Suppl.1):S15–S27.
survival programmes in Africa: Combatting Childhood
Communicable Diseases Project. Int J Epidemiol 1993; 22(Suppl.1):
S8–14.
6 Adetunji J. Trends in under-5 mortality rates and the HIV/AIDS
3 Ahmad OB, Lopez AD, Inoue M. The decline in child mortality: a
7 Walker N, Schwartlander B, Bryce J. Meeting international goals in
reappraisal. Bull World Health Organ 2000;78:1175–91.
epidemic. Bull World Health Organ 2000;78:1200–06.
child survival and HIV/AIDS. Lancet 2002;360:284–89.
MONITORING OF CHILD MORTALITY
8 Millennium Assembly of the United Nations. Millennium Development
Goals. 2000. http://www.developmentgoals.org/Child_Mortality.htm
9 The Global Fund against AIDS, TB and malaria. http://www.
globalfundatm.org/
10 World Health Organization. World Health Report 2002. Reducing Risks,
Promoting Healthy Life. Geneva: World Health Organization. no. WHA
540.1, 2002.
1301
22 Lengeler C. Insecticide-treated bednets and curtains for preventing
malaria. Cochrane Database Syst Rev 2000; CD000363.
23 Phillips-Howard PA, Nahlen BL, Kolczak MS et al. Efficacy of
permethrin-treated bednets in the prevention of mortality in young
children in an area of high perennial malaria transmission in western
Kenya. Am J Trop Med Hyg 2003;68:23–39.
24 Kidane G, Morrow RH. Teaching mothers to provide home treatment of
11 Korenromp EL, Williams BG, Gouws E, Dye C, Snow RW.
malaria in Tigray, Ethiopia: a randomised trial. Lancet 2000;356: 550–55.
Measurement of trends in childhood malaria mortality in Africa: an
assessment of progress toward targets based on verbal autopsy. Lancet
Infect Dis 2003;3:349–58.
25 Byass P, Berhane Y, Emmelin A et al. The role of demographic
12 Black RE, Morris SS, Bryce J. Where and why are 10 million children
26 Rutstein SO. Fertility levels, trends, and differentials 1995–1999. 2002:
dying every year? Lancet 2003;361:2226–34.
December. ORC Macro, Calverton MD. DHS Comparative Reports no. 3.
13 Roll Back Malaria. Framework for Monitoring Progress and Evaluating
Outcomes and Impact. Geneva: World Health Organization. no.
WHO/CDS/RBM/2000.25, 2000.
14 Mung’ala VO, Snow RW. Death registration on the Kenyan Coast. East
Afr Med J 1994;71:747–50.
15 ORC Macro, MEASURE DHS+. Demographic and Health Surveys
(DHS). ORC Macro,
measuredhs. com
Calverton,
MD,
surveillance systems (DSS) in assessing the health of communities: an
example from rural Ethiopia. Public Health 2002;116:145–50.
USA.
http://www.
16 Hill AG, David PH. Monitoring changes in child mortality: new
27 Curtis S. Assessment of the quality of data used for direct estimation
of infant and child mortality in DHS-II surveys. 1995: March. Macro
International Inc., Calverton MD. Occasional Papers no. 3.
28 Ryder RW, Nsuami M, Nsa W et al. Mortality in HIV-1-seropositive
women, their spouses and their newly born children during 36
months of follow-up in Kinshasa, Zaire. AIDS 1994;8:667–72.
29 McDermott JM, Slutsker L, Steketee RW, Wirima JJ, Breman JG,
Heymann DL. Prospective assessment of mortality among a cohort
of pregnant women in rural Malawi. Am J Trop Med Hyg 1996;55:66–70.
methods for use in developing countries. Health Policy Plann
1988;3:214–26.
30 Ng’weshemi J, Urassa M, Isingo R et al. HIV impact on mother and child
17 Bicego G, Ahmad OB. Infant and child mortality. Demographic and
31 Crampin AC, Floyd S, Glynn JR et al. The long-term impact of HIV and
Health Surveys Comparative Studies No. 20. 1996: August. Macro
International Inc., Calverton MD. Demographic and Health Surveys
Comparative studies no.
18 Altman DG, Machin D, Bryant TN, Gardner M (eds). Statistics with
Confidence. 2nd Edn. London: British Medical Journal University
Publications, 2000.
19 Boerma JT, Sommerfelt AE. Demographic and health surveys (DHS):
contributions and limitations. World Health Stat Q 1993;46:222–26.
20 Roll Back Malaria. The African summit on Roll Back Malaria. 2000:
April 25, 2000. World Health Organization, Abuja. no. WHO/
CDS/RBM/2000.17.
21 Jones G, Steketee RW, Black RE, Bhutta ZA, Morris SS. How many
child deaths can we prevent this year? Lancet 2003;362:65–71.
mortality in rural Tanzania. J Acquir Immune Defic Syndr 2003;33:393–404.
orphanhood on the mortality and physical well-being of children in
rural Malawi. AIDS 2003;17:389–97.
32 World
Health Organization Division of Child Health and
Development. Integrated management of childhood illness:
conclusions. Bull World Health Organ 1997;75(Suppl.1):119–28.
33 Hill K, Pande R, Mahy M, Jones G. Trends in child mortality in the
developing world: 1960 to 1996. New York: UNICEF, 1999.
34 UNICEF. Multiple Indicator Cluster Surveys (MICS-2)/End of Decade
Assessment. 2001. http://www.childinfo.org/MICS2/
35 Korenromp EL, Miller J, Cibulskis RE, Cham MK, Alnwick D, Dye C.
Monitoring mosquito net coverage for malaria control in Africa:
possession versus use by children under 5 years. Trop Med Int Health
2003;8:693–703.
Appendix Least-squares linear regression of determinants of standard errors on under-5 mortality estimates (5q0) for the interval of 0–4 years
preceding 41 Demographic and Health Surveys (DHS) surveys conducted in sub-Saharan African countries between 1986 and 2002.
SE = standard error on the coefficient, LN = natural logarithm, R2 = proportion of explained variation.
Parameter
Coefficient (SE)
P-value
7.912 (1.268)
0.001
LN of (5q0/1000)/(1 (5q0/1000))
2.165 (0.266)
0.001
Reciprocal of the square root of the total fertility rate
9.918 (2.751)
0.001
Intercept
Reciprocal of the square root of the number of women interviewed
Design effect
Number of observations: 41, degrees of freedom in model: 3, R2 = 0.935.
583.504 (36.27)
0.001
5.322 (0.344)
0.001