We use data from surveillance of burials and from verbal autopsy interviews. The burial surveillance was initiated at all cemeteries of Addis Ababa in February 2001 and registers around 20,000 deaths a year [14
]. At the end of 2007, the surveillance covered 62 Orthodox Christian, 7 Muslim, 2 Catholic, 1 Jewish, and 9 municipal cemeteries. The largest of the municipal cemeteries, Baytewar
buries persons without close relatives or friends to facilitate a funeral. Baytewar alone accommodates approximately 15% of burials, and most of its records lack socio-demographic background information. Over 40% of persons buried at Baytewar are infants.
The surveillance is conducted by cemetery clerks who are regularly briefed in training workshops. The clerks collect information on the date of burial, age, name, sex, address, marital status, region of birth, ethnicity, religion, and a lay report of the cause of death from relatives or close friends while they make arrangements for burial. Just 5.3% of the records have missing values for age and 5.5% have missing values for sex. Excluding cases from the Baytewar cemetery, fewer than 1% have missing values for either age or sex. Missing values for age and sex are imputed using a hot-deck procedure [22
Infants who die before the naming ceremony (40 days for boys and 80 days for girls) are often not given a formal funeral. The burial surveillance is thus prone to underreporting of infant mortality. After correcting the underreporting of early childhood deaths, the crude death rate (CDR) for 2001 was estimated at just over 9 per 1,000 [14
]. Other potential sources for underreporting are the burial of residents beyond the city administration limits, the return of terminally sick migrants to their families for care [23
], the repatriation of bodies for burial, failure of cemetery clerks to register burials, and possibly also illegal burials.
The verbal autopsies (VAs), our second data source, are post-mortem interviews with close relatives or caretakers of the deceased about the signs and symptoms experienced during terminal illness [25
]. The VAs were administered for adult deaths (twelve years and above) randomly selected from burials registered in November and December of 2003. Most records from the Baytewar cemetery were de facto excluded for a VA interview because selection required complete information for the name, age, sex, and address of the decedent.
VAs were conducted by a pair of trained community health workers who visited the household two to four weeks after the death. Of the cases selected for interview, 78.6% of the interviews were completed, 4.5% of the caretakers refused an interview, 13.8% of the households could not be found, and 2.9% of the interviews were not completed for other reasons. Eleven questionnaires were discarded because interviewers doubted the truthfulness of the respondent’s answers.
Causes of death were ascertained via physician review. Two physicians independently assigned an underlying cause of death. If the assigned ICD10 code (three digits) for the first two physicians did not match, the VA questionnaire was reviewed by a third physician. In eleven cases the cause of death could not be established. The data used in this paper are the remaining 413 cases (ages 20–64) with a physician-assigned cause of death. These were classified as AIDS and non-AIDS deaths. We group TB and AIDS together because they are highly correlated and both laymen and physicians tend to have difficulties distinguishing both causes of death. For validating our estimation method, we also report on the results of 141 and 625 VAs that were conducted in 2001 and 2007, respectively. Only minor changes were implemented in the questionnaires and fieldwork procedures in the different VA survey rounds.
To demonstrate mortality trends, we (1) compare observed age and sex-specific deaths with those implied in population projections excluding an AIDS effect, and (2) develop a method to estimate the number of AIDS deaths from lay reports of causes of death.
Observed versus projected deaths
To estimate the annual number of deaths that would occur in the absence of HIV/AIDS, we use the Spectrum package (v.3.14) [26
] to emulate population projections from the Central Statistical Authority (CSA) [9
]. These use 1994 census population as the baseline and assume a steady increase in life expectancy (e0
) from e0m
: 56.2 and e0f
: 58.8 to e0m
: 62.0 and e0f
: 66.0 between 1995 and 2007. We assume a decline in the Total Fertility Rate from 2.14 in 1995 to 1.94 in 2000 and 1.80 in 2007. These values reflect that fertility in Addis Ababa is below replacement [27
]. For migration, we follow CSA’s high variant assumption of a constant number of in-migrants at the 1994 level. With the exception of the implied effect of AIDS on the population age structure in the 1994 census, these projections thus ignore the demographic impact of AIDS. We compare the projected deaths (by age and sex) with observed numbers from the burial surveillance.
Extrapolations from lay reports of causes of death
The second, less common, method for estimating mortality trends is based on lay reports of causes of death. shows the distribution of lay reports for 2002 (first full year of observation) and 2007. Although few deaths are explicitly ascribed to HIV/AIDS, some of the lay diagnoses are very suggestive of AIDS. This is also demonstrated in , which maps lay reports against the cause of death attributed by physicians in the sample with a reviewed verbal autopsy. Deaths explicitly labeled as TB or AIDS deaths have a positive predictive value (PPV) of 89%, but account for only 7% of AIDS deaths. Herpes zoster, diarrhea, uterine cancer, mental problem, and emaciation account for another 8% of AIDS deaths, and their PPV is 71%. Lung disease and cough are other good predictors of AIDS mortality, and about 40% of all adult cases are labeled as such. Cold accounts for just under 25% of all AIDS deaths. Together these lay reports accurately predict an AIDS death 80% of the time and have a sensitivity (SE) of 78%. Their cumulative specificity (SP) is 77%. Henceforth, this group of lay reports is labeled “AIDS-indicative lay reports” (LRA). Other lay reports of causes of death are less predictive of AIDS ascriptions by physicians (PPV: 25%, SE: 22%, SP: 23%).
Distribution of lay reports of causes of death by sex (Addis Ababa, age 20–64, 2002 and 2007)
Diagnostic values of lay reports of causes of death for identifying AIDS mortality (Addis Ababa, both sexes, age 20–64, 2003–4)
To estimate trends in AIDS mortality, we use the PPV and SE of the AIDS-indicative lay reports as anchors. This method was first explored in Araya et al. [20
], and is extended to allow for age (x
) and sex-specific variation in the diagnostic validity of the lay reports. In the subsample with a verbal autopsy diagnosis of the cause of death, the PPV and SE of lay reports are estimated by probit regression models (only age has a significant–negative–effect on the PPV). We then estimate the year and sex-specific number of AIDS deaths (DA
) in the age range 20–64 as:
DA is first estimated for deaths that come with a sufficiently specific lay report (categories 1–5 in ). The distribution of causes of death in deaths with an unknown or weakly specified lay report (category 6 in ) is assumed to be the same as for other deaths and is added to the total of DA. Confidence intervals around DA are estimated by bootstrapping. These account for the uncertainty in the estimates of the PPV and SE as well as the stochastic variation in the imputation of missing values for age and sex.
This approach presumes that the diagnostic validity of lay reports remains stable over time, and that is not necessarily realistic. However, comparison of AIDS mortality fractions estimated by the method described above and those obtained from VAs conducted in 2001 and 2007 is encouraging: extrapolations from the lay reports establish the AIDS mortality fraction in the age range 20–64 in 2001 at 59.4% and 67.0%, and for 2007 at 45.4% and 48.8% for men and women, respectively. The VA estimates for 2001 are 58.4% (N=77) and 67.2% (N=64), and for 2007 they are 44.9% (N=314) and 51.1% (N=311) for men and women, respectively.
To determine the number of averted AIDS deaths, we compare the estimated number of AIDS deaths with the implied number of AIDS deaths in population projections that incorporate the effect of HIV/AIDS, but do not account for the impact of ART. These projections use the baseline population and projection parameters described earlier, and make additional assumptions about the course of the HIV/AIDS epidemic. As before, the projections are done with Spectrum [26
]. Estimates of HIV prevalence were generated using the UNAIDS Estimation and Projection Package (EPP 2007) [29
] with inputs from ANC data [13
] and population-based seroprevalence surveys done in 1994 [30
] and 2005 [12
]. With these inputs, HIV prevalence is estimated to have reached 1.0% in the mid 1980s, peaked at 7.9% in 1997, and declined to 7.4% by 2007. The ratio of female to male infections is assumed to have increased from 1.00 to 1.35 between 1994 [30
] and 2007.