Estimates of maternal mortality for 1995.

Abstract

To present estimates of maternal mortality in 188 countries, areas, and territories for 1995 using methodologies that attempt to improve comparability.
For countries having data directly relevant to the measurement of maternal mortality, a variety of adjustment procedures can be applied depending on the nature of the data used. Estimates for countries lacking relevant data may be made using a statistical model fitted to the information from countries that have data judged to be of good quality. Rather than estimate the Maternal Mortality Ratio (MMRatio) directly, this model estimates the proportion of deaths of women of reproductive age that are due to maternal causes. Estimates of the number of maternal deaths are then obtained by applying this proportion to the best available figure of the total number of deaths among women of reproductive age.
On the basis of this exercise, we have obtained a global estimate of 515,000 maternal deaths in 1995, with a worldwide MMRatio of 397 per 100,000 live births. The differences, by region, were very great, with over half (273,000 maternal deaths) occurring in Africa (MMRatio: > 1000 per 100,000), compared with a total of only 2000 maternal deaths in Europe (MMRatio: 28 per 100,000). Lower and upper uncertainty bounds were also estimated, on the basis of which the global MMRatio was unlikely to be less than 234 or more than 635 per 100,000 live births. These uncertainty bounds and those of national estimates are so wide that comparisons between countries must be made with caution, and no valid conclusions can be drawn about trends over a period of time.
The MMRatio is thus an imperfect indicator of reproductive health because it is hard to measure precisely. It is preferable to use process indicators for comparing reproductive health between countries or across time periods, and for monitoring and evaluation purposes.

Full text

Page 1

Research
Estimates of maternal mortality for 1995
Kenneth Hill,1Carla AbouZahr,2& Tessa Wardlaw3
Objective To present estimates of maternal mortality in 188 countries, areas, and territories for 1995 using
methodologies that attempt to improve comparability.
Methods For countries having data directly relevant to the measurement of maternal mortality, a variety of
adjustment procedures can be applied depending on the nature of the data used. Estimates for countries lacking
relevantdatamaybemadeusingastatisticalmodelfittedtotheinformationfromcountriesthathavedatajudged
to be of good quality. Rather than estimate the Maternal Mortality Ratio (MMRatio) directly, this model estimates
theproportionofdeathsofwomenofreproductiveagethatareduetomaternalcauses.Estimatesofthenumberof
maternal deaths are then obtained by applying this proportion to the best available figure of the total number of
deaths among women of reproductive age.
Findings Onthebasisofthisexercise,wehaveobtainedaglobalestimateof515000maternaldeathsin1995,with
a worldwide MMRatio of 397 per 100000 live births. The differences, by region, were very great, with over half
(273000 maternal deaths) occurring in Africa (MMRatio: >1000 per 100000), compared with a total of only
2000 maternal deaths in Europe (MMRatio: 28 per 100000). Lower and upper uncertainty bounds were also
estimated, on the basis of which the global MMRatio was unlikely to be less than 234 or more than 635 per
100000 livebirths.Theseuncertaintyboundsandthoseofnationalestimatesaresowidethatcomparisonsbetween
countries must be made with caution, and no valid conclusions can be drawn about trends over a period of time.
Conclusion The MMRatio is thus an imperfect indicator of reproductive health because it is hard to measure
precisely.Itispreferabletouseprocessindicatorsforcomparingreproductivehealthbetweencountriesoracrosstime
periods, and for monitoring and evaluation purposes.
Keywords: Maternal mortality/statistics; Statistics/methods; Models/statistical (source: MeSH).
Mots cle ´s: Mortalite ´ maternelle/statistique; Statistique/me ´thodes; Mode `le statistique (source: INSERM).
Palabras clave: Mortalidad materna/estadı´stica; Estadı´stica/me ´todos; Modelos estadı´sticos (fuente: BIREME).
Bulletin of the World Health Organization, 2001, 79: 182–193.
Voir page 192 le re ´sume ´ en franc ¸ais. En la pa ´gina 193 figura un resumen en espan ˜ol.
Introduction
Maternal mortality is notoriously difficult to measure
(1). The most widely used measure, the Maternal
MortalityRatio(MMRatio),expressesmaternaldeaths
per 100000 live births, but MMRatios rarely exceed
1000 or 1 per 100 live births. Maternal deaths are thus
relatively rare events and they are also hard to identify
precisely, both of which limit the applicability of
sample survey measurement methods. In countries
with well-developed statistical services, the conven-
tional source of information about maternal mortality
isthecivil registration system, whichrecords bothlive
births and deaths, by cause, on a continuous basis.
Even in such settings, however, maternal deaths are
invariably found to be under-recorded in official
statistics owing to misclassification of the cause of
death (2–4). In countries with less well developed
statistical services, the outright omission of deaths
contributes an additional source of error.
Despite the difficulties in measuring maternal
mortality, interest in obtaining estimates has in-
creased. The MMRatio varies by a factor of over
100 between the highest and lowest mortality
settings, making it the health outcome with the
largest gap between developed and developing
countries. Interest has also increased because one
of the main targets of the Programme of Action,
whichwasdevelopedattheInternationalConference
onPopulationandDevelopmentinCairoin1994(5),
is reproductive health. One specific goal of the
1Professor and Director, Hopkins Population Center, Johns Hopkins
University, School of Hygiene and Public Health, 615 North Wolfe
Street, Baltimore MD 21205-2179, USA. Correspondence should
be addressed to this author.
2Coordinator,Policy,PlanningandEvaluation,FamilyandCommunity
Health, World Health Organization, Geneva, Switzerland.
3Senior Project Officer, Statistics and Monitoring, Division of
Evaluation, Policy and Planning, UNICEF, New York, USA.
Ref. No. 00-0865
182
# World Health Organization 2001Bulletin of the World Health Organization, 2001, 79 (3)

Page 2

Programme of Action called for quantified reduc-
tions in maternal mortality. However, although the
goal was phrased in terms of a 1990 baseline, no
global estimates of the situation in 1990 were
available. To try to bridge the gap between the need
for reliable data and the quality of conventional data
sources,alternativeapproachestomeasurementhave
been explored.
In1996,WHOandUNICEFpreparedmodel-
based estimates of maternal mortality for about the
year 1990 in developing countries, for which
adequate empirical information had not been
identified (6, 7). The model caused considerable
controversy, particularly in countries which did have
relevant information and for which the model-based
estimates were often substantially higher than the
existing figures. The resulting controversy, however,
had at least three positive outcomes: 1) it increased
general awareness of the problems surrounding the
measurementoftheMMRatio;2)itdrewattentionto
thevariousstrengthsandweaknessesofthedifferent
measurement approaches; and 3) it brought to light
relevant data hitherto unavailable to the international
community, or stimulated new data collection, or
both. This paper updates the 1996 exercise using a
revised methodology, incorporating new data as
available, and presents the maternal mortality
estimates, by countries and regions, for 1995.
The basis of the estimates
Theestimationstrategywasdevelopedinresponseto
concerns in three major areas, which were expressed
in international forums about the 1996 exercise:
1) failure to use country data where they were
availableandwhicharebelievedtobeofgoodquality;
2) concern about the choice of a dependent variable
in the statistical model; and 3) rigidities in the model
that limited the range of possible values of estimates
of maternal mortality (8–11).
For the present estimation exercise, every
attemptwasmadetousecountrydatawhereavailable
and of adequate quality. Using the available data,
countries, areas, and territories were placed in one of
the following six categories, according to whether
they had:
(a) complete death registration data, with generally
good cause of death attribution;
(b) complete death registration data, but uncertain
cause of death attribution;
(c) information from a survey on deaths of sisters,
with timing of death related to pregnancy;
(d) recentresultsfromaReproductiveAgeMortality
Study (RAMOS);
(e) other available relevant data;
(f) no relevant data available for a recent period.
Different strategies were used to arrive at estimates
for the countries in each category, taking advantage
ofthestrongerelementsofthedatawhileminimizing
the distortions resulting from data errors.
Despite increased data availability, estimates
for countries in categories (b) and (f) still require a
statistical model to estimate the MMRatio. The
model isfittedtocountryobservations deemedtobe
of very high quality,using predictor variables that are
availableforalmosteverycountry,andisthenusedto
predict estimates for countries that lack high quality
information.
Choice of a dependent variable for the
statistical model
Amodelusingtheproportionofdeathsofwomenof
reproductive age due to maternal causes (PMDF),
rather than the MMRatio, has several advantages.
First, the estimated PMDF is boundedby 0 and 1, so
that its logit (ln{PMDF/(1-PMDF)}) can be mod-
elled without risk of the predicted values falling
outside that range. The predicted PMDF can then be
applied to an ‘‘envelope’’ of deaths of women of
reproductive age, obtained either from death regis-
tration data or from special tabulations of deaths, by
age and sex, provided by the United Nations
PopulationDivision(UNPD)fromthe1998revision
of World Population Prospects (12), to estimate the
number ofmaternaldeaths.Thus,theestimates have
tobeconsistentwithotherdemographicinformation
aboutthepopulationinquestion.Second,thePMDF
estimate makes full use of limited data. For example,
in category (b) countries with good overall death
registrationbutuncertain cause-of-deathreporting, a
model-basedestimateofPMDFcanbeappliedtothe
‘‘envelope’’ of registered female deaths, modelling
only the distribution of such deaths by cause, not
their number.
Third, sisterhood data, the most frequently
available type of information about maternal mortal-
ity in high-mortality settings, are likely to provide
more robust measures of the PMDF than of the
MMRatio. A large majority of nationally representa-
tive data sets on sister survival, mostly collected
under the auspices of the Demographic and Health
Surveys(DHS)programme,haveusedsiblinghistory
to collect the basic data. Each of a respondent’s
siblings is listed, and survival status is recorded. For
sisters who died at age 12 or older, further questions
are asked as to whether she was pregnant, or giving
birth,orwithintwomonthsoftheendofapregnancy
when she died. These questions are used to identify
possiblematernaldeaths;strictlyspeaking,itisatime-
of-death (pregnancy-related) definition, not a cause-
of-death (maternal) definition. Some non-maternal
deaths are included, but some true maternal deaths
arelikelytobeexcluded because therespondent may
not be aware of a sister’s pregnancy at the time of
death. The extent to which these two errors cancel
out is not known with certainty (13, 14), but it is
unlikely that the net error is large. Consequently, the
PMDF can be expected to be estimated with
reasonable accuracy.
Sisterhood estimates of the MMRatio, on the
other hand, depend both on the classification of
183
Bulletin of the World Health Organization, 2001, 79 (3)
Maternal mortality estimates for 1995

Page 3

deaths as pregnancy related or otherwise and on the
completeness of reporting of the sisters’ deaths.
Certain features ofthedata raisequestions about this
completeness. The sibling history format allows the
calculationofmortalityratesfordefinedtimeperiods
before the survey. In an analysis of 15 DHS data sets
(15), the estimates of female mortality between the
ages of 15 and 50 years were higher for the period
0–6 years before the survey than for the period
7–13 years before the survey in a large majority of
countries. These increases were not paralleled by
increases in mortality in other age ranges, such as
under the age of 5 years. Thus the sisterhood data
indicatedimplausible recent increases inadultfemale
mortality in most of the countries studied. Despite
these differences in mortality level, the PMDF
remained remarkably stable across time periods. A
more likely explanation than rising mortality is data
error, either an omission of sister deaths that
occurred further back in the past, or misdating of
suchdeathsandtransferring theminto amore recent
period.Eitherway,onecanhavelittleconfidencethat
the level of mortality for the period 0–6 years before
the survey is accurate. This conclusion is reinforced
by a comparison of sisterhood estimates with
estimates from other sources. For most countries,
there are no alternative sources of estimates, but in
four cases (Guatemala, the Philippines, Senegal and
Zimbabwe) it has been possible to compare the
sisterhood estimates of adult mortality with inde-
pendently validated alternative estimates. In all four,
sisterhood data underestimated recent adult mortal-
ity, by amounts ranging from 15% to 60% (16).
These findings suggest that sisterhood data
tend to underestimate overall mortality, particularly
for periods further back in the past, and that, in the
absence of counterbalancing errors, the MMRatios
from sisterhood estimates are likely to be too low.
The nature of these possible biases in the sisterhood
estimates argue for using such data in the form of
PMDFs rather than MMRatios.
One technical problem has to be addressed
before using the sisterhood PMDFs. The PMDF
shown in DHS country reports is calculated as the
number of pregnancy-related deaths of sisters
divided by the overall number of sister deaths.
However, the distributions of sister deaths and of
sister-yearsofexposurebyagearenotthesameasthe
corresponding distributions in the actual population
(17). For example, the sisters of reproductive age of
respondents aged 15–19 years are likely to be, on
average, older than the respondents (they cannot be
younger than 15, but they can be 20 or older),
whereas the sisters of reproductive age of respon-
dents aged 45–49 years are likely to be generally
younger. Years of exposure of sisters are thus
concentrated in the central ages of the reproductive
period at the expense of the extremes. However, it is
alsointhecentralagesthatmostbirths,andtherefore
most maternal deaths, are likely to occur. Thus, the
reported PMDF is likely to be higher than the true
PMDFwouldbeforagroupofwomendistributedby
ageinthesamewayastheactualpopulation.Inorder
to allow for this effect, age-standardized PMDFs
were calculated, which can appropriately be applied
to the number of deaths of women of reproductive
age to estimate the number of maternal deaths.
The use of information from RAMOS studies
also argues for the use of the PMDF rather than the
MMRatio in the modelling exercise. RAMOS studies
typically use a variety of data sources to try to get as
complete a count of the number of deaths as
possible, and of the maternal deaths among the
deaths. However, RAMOS studies do not typically
go to great lengths to check the numbers of births to
use as the denominator of the MMRatio. Thus
RAMOS studies, even though they tend to report
their results in terms of the MMRatio, are by design
more likely to obtain unbiased estimates of the
PMDF.
Choice of independent variables
Thechoiceofindependentvariablesisconstrainedby
theneedtouseonlyvariablesthatareavailableforthe
vast majority of countries in the world around the
year 1995. Given the nature of the dependent
variable, the first predictor variable that has to be
included is an indicator of fertility: the higher fertility
(for a given maternal mortality risk per birth), the
higherthenumberofmaternaldeaths,andthehigher
the PMDF. Since PMDF is calculated on the basis of
the age distribution of women aged 15–49 years, the
appropriate fertility measure is the General Fertility
Rate (GFR).
Socioeconomic variables are constrained by
theirlimitedavailability,butitwaspossibletoinclude
both female literacy and per capita income in
purchasing power parity terms. Neither variable
retained significance, however, once other variables
were included, although colinearity was high.
As an indicator of the role of health services,
TRATT, the percentage of deliveries assisted by a
skilled attendant — i.e. a physician, nurse, or
professionally-trained midwife) was included in the
model. Given the widespread consensus that
maternal mortality will not be significantly reduced
if the appropriate management of obstetric compli-
cations is absent, this variable comes closest to
capturing such access while being widely available
from DHS and other surveys, and compiled into a
database by WHO and UNICEF.
The HIV/AIDS epidemic can be expected to
have important effects. A major increase in female
adult mortality resulting from the epidemic would
tend to reduce the PMDF by increasing the other
causes of death (and perhaps by reducing fertility).
Failure totakethisepidemicintoaccountwouldthus
result in a model that would overestimate PMDF in
countries badly affected by the epidemic, and would
tend to underestimate the PMDF elsewhere. Coun-
try-specific estimates of HIV prevalence for 1995
(HIVAIDS)from UNAIDS were therefore incorpo-
rated in the model.
184
Bulletin of the World Health Organization, 2001, 79 (3)
Research

Page 4

A variety of variables reflecting data quality
were tried, but the variable ultimately used was a
dummy variable identifying a country with death
registration reported to the United Nations to be
complete(goodVR).Twodummyvariablesforregion,
oneidentifyingcountriesofformerlysocialistEurope
(FSE) and the other identifying countries of Latin
America, Africa, West and South Asia (LASSAME)
were also included.
Selection of cases for model fitting
National estimates of maternal mortality were care-
fully reviewed on a case-by-case basis, and only
adequately documented estimates, backed by clear
descriptions of acceptable methodology, were in-
cluded in the data set to which the model was fitted.
In some cases, adjustments were made to the
dependent variable before fitting the model. For
countries with documented assessments of comple-
teness of recording of maternal deaths (Argentina,
Costa Rica, Finland, France, Mexico, the Nether-
lands, New Zealand, and the United Kingdom), that
estimate of the coverage was used to inflate the
observed values. PMDFs for other statistically
developed countries with registration-based esti-
mates were inflated by a uniform factor of 1.5a(7).
Estimates of PMDF derived from sisterhood data
were age-standardized as noted above. Estimates
derived from RAMOS studies were not adjusted.
Independent variables were also carefully reviewed
where possible. In particular, estimates of the
proportionofdeliveries assisted bymedically-trained
professionals were reviewed country by country.
Inall,themodelwasfittedto73contemporary
observations, 30 of which were for industrialized
countries or countries of formerly socialist Europe.
Thedataset,excludingvariablesthatdidnotproveto
be significant, is shown in Annex Table 1 (available
on our web site: http://www.who.int/bulletin).
The final model
The final model, fitted using a robust regression
approach that underweights the observations with
large residuals,bis shown below. Values in parenth-
eses under the coefficients show the t-values; all
except the HIVAIDS variable were significant at the
1% level or better. The HIVAIDS variable was close
to statistical significance at the 10% level, and its
coefficient had the expected sign. Including this
variable in the model was judged the most appro-
priate way of avoiding possible prediction bias that
might result from inflated numbers of non-maternal
HIV-related deaths The robust regression does not
provideadirectR2forthismodel,butOrdinaryLeast
Squares regression gave very similar parameter
estimates with an R2of 0.919 and a Root Mean
Square Error of 0.51. A plot of predicted values of
PMDF plotted against observed values (not shown)
does not indicate departure from the standard
assumptions of regression analysis.
Estimates of maternal mortality
for 1995
Maternal mortality estimates for countries with a
population of 300000 or more are shown as
MMRatios in Tables 1(a) to 1(f) on the basis of data
source, following the typology described earlier. The
ways in which final values were arrived at vary from
one country to another according to the country’s
data availability category, as follows.
. Forthe48countriesincategory(a),theMMRatios
were based on the reported values adjusted by a
factor that wasinmost cases1.5.Theseestimates,
withtheactualadjustmentfactorsused,areshown
in Table 1(a).
. For the 18 countries in category (b) — based on
complete death registration but questionable
cause of death data — the PMDF was estimated
from the model, and applied to the appropriate
‘‘envelope’’ of female deaths to estimate the
number of maternal deaths; the MMRatio was
obtained by dividing by the registered number of
live births. The results for these countries are
shown in Table 1(b).
. The 28 countries in category (c) — based on direct
sisterhood estimates — are listed in Table 1(c).
Each age-standardized PMDF from the sisterhood
data was applied to the number of deaths among
females aged 15–49 years, indicated in the United
NationsEstimatesandProjections(1998Revision)
(18) for the year 1995. There were two exceptions
to this treatment. For Zimbabwe, a review of data
on household deaths by age from the 1992
Population Census suggested that the data were
of good quality; the sisterhood PMDF was there-
foreappliedtotheCensusestimateofthenumbers
of deaths of women of reproductive age. The
second exception concerns countries with sister-
hood estimates and with substantial AIDS mortal-
ity. The sisterhood PMDFs apply to a time period
earlier than 1995. For the United Republic of
Tanzania, for example, the PMDF covers a period
from1987to1996.Thenumberoffemaledeathsin
this country rose steeply in the early 1990s as a
aThe value of 1.5 was arrived at by averaging the results of studies
of underreporting of maternal deaths in France, the United Kingdom,
and USA.
bThe rreg command in STATA was used.
ln {PMDF / (1–PMDF)} = – 8.289 – 0.0141*TRATT + 1.386*lnGFR + 0.682*FSE + 0.719*LASSAME – 0.684*goodVR – 0.0197*HIVAIDS
(–4.54)(–2.96)(3.91) (3.19)(2.93) (–3.16)(–1.36)
185
Bulletin of the World Health Organization, 2001, 79 (3)
Maternal mortality estimates for 1995

Page 5

resultoftheAIDSepidemic.Applyingthe1987–96
PMDF to the 1995 deaths would therefore over-
estimate the number of maternal deaths. For all
countries with substantial numbers of AIDS
deaths,theUnitedNations estimates ofnon-AIDS
and AIDS deaths for 1985–90 and 1990–95 were
used to estimate the proportion, p, of all deaths of
femalesofreproductive agethatwereduetoAIDS
Table 1(a). MMRatios in 48 countries with good registration systems and relatively good attribution of
cause of death
CountryYearReported
MMRatio
Adjustment
factor
Adjusted
MMRatio
No. of live births
(Demographic
Yearbook)
Argentinaa
Australia
Austria
Belarus
Belgium
Bosnia and Herzegovina
Bulgaria
Canada
Costa Ricaa
Croatia
Cyprus
Czech Republic
Denmark
Estonia
Finlanda
Francea
Germany
Greece
Hungary
Ireland
Israel
Italy
Japan
Latvia
Lithuania
Luxembourg
Malta
Mauritius
Netherlandsa
New Zealanda
Norway
Poland
Portugal
Puerto Rico
Republic of Moldova
Romania
Russian Federation
Singapore
Slovakia
Slovenia
Spain
Sweden
Switzerland
The former Yugoslav Republic
of Macedonia
Ukraine
United Kingdoma
USA
Yugoslavia
1995
1993
1995
1996
1989
1990
1992–94
1993
1994
1995
1993
1995
1995
1995
1993–95
1992–94
1994–95
1993–95
1995
1990–92
1992–95
1990–92
1992–94
1994
1996
1994
1994
1996
1993–95
1994
1991–93
1994–96
1993–95
1991
1996
1997
1996
1993–95
1995
1995–96
1990–92
1993–95
1993–94
1995–96
441.9
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.2
1.5
1.5
1.5
1.5
1.5
1.03
2.0
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.4
1.0
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
1.5
84 658 735
256 190
88 669
101 144
115 638
43 415b
71 967
378 011
80 306
50 182
9869
96 097
69 771
13 560
63 067
729 609
765 221
101 495
112 054
48 530
116 461
526 064
1 187 064
21 595
41 180
5421
5003
20 604
190 513
57 795
60 292
433 109
107 184
64 325c
56 411
236 640
1 363 806
48 635
61 427
18 980
363 469
103 326
82 203
32 154
4
7
6
11
33 22
58
10
15
15
23
46
29
12
35
18
0
9
0
14
15
78
10
52
66
1020
128
12
1523
6
5
7
8
9
8
11
12
68
27
45
18
0
0
0
0
3045
10
15
7
15
6
8
8
9
12
12
30
63
62
74
20
42
41
49
6
9
9
14
1711
5.5
5
5
11
8
8
8
17
1996
1992–95
1990–95
1995–96
30 1.5
1.4
1.5
1.5
45
10
12
15
492 861
732 049
3 899 589
140 504
7
8
9.7
aNational adjustment factors.
bBirths from UN Population Division estimates.
c1994 births.
186
Bulletin of the World Health Organization, 2001, 79 (3)
Research