Sampling Design

DRC was divided into 2 strata along the 2001 military front line: an east stratum of territory formerly held by rebel groups and a west stratum formerly held by government forces. We surveyed each stratum using a 3-stage, household-based cluster sampling technique.

Sample size was calculated to detect a difference between east and west assuming the west had reverted to the sub-Saharan baseline rate for 2005 (1.4 deaths per 1000 population per month)8 and the east had reduced to the 2003–2004 survey result for the west (1.8).Reference Coghlan, Brennan and Ngoy5 We used design effects from the 2004 IRC survey for the east (5.3) and west (3.4). We used Optimal Design for Multi-level and Longitudinal Research version 1.77 (University of Michigan, Ann Arbor) to then determine the optimal number of health zones, clusters per health zone, and households per cluster, following consideration of the survey budget, logistical issues, the intraclass correlations, and average household sizes from our 2004 survey. In the east, the intraclass correlation at the health zone level was 0.51 and the average household size was 7.2. These figures were 0.18 and 5.2, respectively, in the west. Ethical approval for the study was provided by the ethics committee of the School of Public Health, Kinshasa University.

In the first stage, 4 east health zones (Kalemie, Kalima, Katana, and Kisangani-Ville) surveyed by the IRC on at least 3 prior occasions were purposely selected to allow for historical comparisons. These were excluded from the sampling frame. A single health zone in the west was excluded because it contained a military base and access was not granted by the Congolese government. From the remaining population, 15 health zones were randomly selected from 261 western zones and 16 health zones were randomly selected from 248 eastern health zones for the study. The probability of selection for an individual health zone was proportional to its population size, using 2006 data from the Congolese Ministry of Health (MOH).9 For example, a health zone with a population of 200,000 had twice the chance of selection as a health zone with a population of 100,000. Total population for DRC was estimated by the MOH at 69.9 million (east stratum 32.6 million; west stratum 37.3 million; inaccessible 44,000).

In the second stage, we selected 20 of the smallest population units within each health zone (villages in rural settings or avenues in urban settings) with a probability of selection proportional to population size. Where populations were unknown, the relative size of smallest units was crudely weighted using estimates of local leaders.

In the third stage, 20 households in each of these smallest units were surveyed. For well-ordered (grid pattern or along a single road) and/or small units (<100 houses), we used systematic random sampling to select households. For widely distributed and/or larger units (>100 houses), households were selected according to the standard World Health Organization Expanded Program on Immunization (WHO/EPI) cluster sampling random walk method. This method entails randomly selecting a direction from the center of the cluster by spinning a pen; counting the houses along that direction to the boundary of the cluster; choosing the first household at random from among those counted; and choosing subsequent households by a rule of proximity (ie, the next closest household). This method was initially developed for estimating vaccination coverage, but it has been used widely to estimate other indicators, including mortality. It is described in more detail elsewhere.10

If occupants could not be found or if they refused to participate, or if no household member older than 16 years was home, then that household was skipped and the next nearest visited. Logistical, security, and time constraints prevented a revisiting of empty households.

Household Data Collection

The survey questionnaire consisted of 3 questions used in the 2002 and 2004 surveys. It was back-translated into French and the 4 most common local languages (Lingala, Kikongo, Tshiluba, and Swahili). The purpose of the study was explained to all heads of households and oral consent was obtained. Wherever possible, the questionnaire was conducted with a senior female member of the household, because our previous experience had indicated that they provided more rapid and accurate answers.

The current household census method was usedReference Checchi and Roberts11 to limit time in remote and insecure sites. The age and sex of people sleeping in the household on the night preceding the interview were documented. All pregnancies, births, or deaths occurring in the household during the recall period (January 1, 2006–April 30, 2007) were recorded. Decedents needed to have resided with the interviewed family at the time of their death or else normally slept in the interviewed household if they died in a health care facility. The age, sex, date of death, and cause of death were recorded for each decedent. We used a predefined list for cause of death that was based on previous IRC mortality survey results and a knowledge of the local epidemiology, with the option of specifying the cause if it did not meet any of our classifications. No independent confirmation of death or verbal autopsy was conducted.

Interviewers were experienced local nursing staff drawn from IRC programs and health zone personnel assigned to the survey by MOH offices. The interviewers received standardized training that included field exercises. Data collection in all health zones was supervised by a senior staff member. When possible, data entry was done in the field. We used personal digital assistants (Dell Axim X30) with Pocket PC Creations version 4.0. All of the data were double entered.

Statistical Analysis

EpiData 3.0 was used to cross-check duplicate entries and errors were compared against original paper questionnaires. STATA 9.2 (StataCorp, College Station, TX) was used for data analysis. The first stage of sampling defined the analysis and the multistage sampling design was accounted for by STATA in the calculation of design effects. Data were weighted according to the probability of selecting each individual in the sample. Mortality rates have been expressed as deaths per 1000 population per month using the equations listed in Table 1. No adjustment is made for live births in the denominator for the under-5 mortality rate (U5MR) because we assumed that the total number of children born during the recall period was equal to the number of children turning 5 during the recall period. Rate ratios were estimated using Poisson regression. We accounted for the survey cluster design in the calculation of all rates and rate ratios.

TABLE 1 Summary of Key Equations

As originally investigated in the 2002 IRC survey, we also analyzed 2 subdivisions within the east stratum reflecting areas with ongoing unrest (designated “East 2002” with 15 health zones) and areas with minimal security concerns since the cessation of conflict (called “Transition East” with 5 health zones). The western division investigated this year is termed “west.” However, the western Congo investigated in the 2002 IRC survey also included the Transition East and for the purposes of this report has been called “West 2002.” Figure 1 illustrates these subdivisions. Figure 2 shows the location of all 35 selected health zones.

FIGURE 1 Stratification of Democratic Republic of Congo for 2007 survey.

FIGURE 2 Location of surveyed health zones with crude mortality rates for 2006–2007.

Mortality Rates

The crude mortality rate in DRC remains elevated at a level that is almost 70% higher than that reported in the 1984 national census6 and more than 55% higher than the reported baseline for sub-Saharan Africa (1.4).8 There has been no reduction in the national crude mortality rate since 2004. As with our previous studies, most deaths have been attributed to the so-called indirect consequences of conflict, especially increased rates of infectious diseases and malnutrition. Children continue to bear a disproportionate burden of the humanitarian crisis in DRC, accounting for almost half of all deaths.

As documented in previous IRC surveys, mortality rates remain significantly higher in the east than in the west.Reference Roberts1–Reference Coghlan, Brennan and Ngoy5 Nonetheless, we found rates that exceeded important international benchmarks on both sides of the former front line. Ngandanjika health zone of Kasai Oriental Province in the west recorded a CMR that was, at best, above the threshold that defines a humanitarian emergency of 0.9 deaths per 10,000 per day (2.7 deaths per 1000 per month; Fig. 2).12 In the east, northern Katanga, known as “Congo’s forgotten crisis,”13 is of particular concern: the CMR in Ankoro health zone was statistically significantly higher than the threshold for a severe emergency of 2.0 deaths per 10,000 per day (6.0 deaths per 1000 per month)14; and Kalemie has had an average death rate above the emergency threshold for every period measured by IRC back to 2000.

Although the national CMR has remained unchanged since 2004, the CMR in the unstable East 2002 region has decreased since the last survey. Security has improved during this period, through a combination of political, diplomatic, and peacekeeping efforts.7, 15 When considered in light of the documented reduction in violent deaths in the present survey and the demonstrated association between insecurity and all-cause mortality from our 2004 study,Reference Coghlan, Brennan and Ngoy5 it is plausible that these improvements have contributed to the decline in CMR in East 2002. Before this survey, all IRC surveys excluded large numbers of eastern health zones due to insecurity,Reference Roberts1–Reference Coghlan, Brennan and Ngoy5 which may have led to past underestimates of the true CMR in East 2002. Consequently, the overall level of decline between our surveys of 2004 and 2007 may have been greater than we have been able to demonstrate.

Perhaps surprisingly, the Transition East recorded significantly increased mortality rates since the last survey. Although this increase in CMR was unexpected, a recent review of mortality surveys in DRC provides some supportive evidence. Two surveys in Equateur in 2006 indicated an emergency situation based on crude mortality, under-5 mortality, and the prevalence of acute malnutrition, and the average CMR in Equateur had worsened since 2004.16

The demographic profile of our sample compared with national records is difficult to interpret because the latter is an adjustment of much earlier census data. Assuming the national data are reasonable, the population deficits we recorded in 2007 are consistent with earlier IRC studies showing markedly elevated mortality among vulnerable children and young males during the intense early years of the conflict.Reference Roberts1–Reference Roberts and Zantop4

Comparisons With Conflicts in Other Countries

Given that there are few comprehensive, longitudinal examinations of the “natural history” of mortality rates following conflict, it is difficult to make confident projections about likely mortality trends in DRC. Nonetheless, other African conflicts may offer relevant comparisons. In 2002, 3 years after the end of the decade-long civil war in Sierra Leone, the average CMR in 4 rural districts—representing one third of the population—was almost 4 times higher than the sub-Saharan baseline.17 In Angola, some provinces documented mortality rates above the emergency threshold 2 years after the end of the war.18 Similarly, Liberia and southern Sudan had sustained increases in mortality years after official peace processes and successful political transitions.19 Finally, a retrospective study of 51 countries demonstrated that civil wars greatly increased the subsequent risk of death and disability for years after the end of conflict.Reference Ghobarah and Russett Huth20

On the one hand, DRC’s experience of sustained increases in mortality is neither unusual compared with other conflicts nor out of keeping with its own history of years of conflict superimposed on decades of socioeconomic and political decline. Indeed, improvements in health indicators in DRC are further constrained by the lowest per capita health expenditures for any country in the world.21 On the other hand, it would be inappropriate to dismiss such sustained elevations of mortality 4 years after the official end of the war as being simply inevitable. For example, a reduction in mortality was achieved in the East 2002 region since 2004, although its CMR is still 85% higher than the regional norm. This is the region of the country that has received the bulk of international peacekeeping support and humanitarian assistance, but at a level that is reported to be out of proportion to the documented need.22, 23 Arguably a more robust and equitable engagement by the Congolese government, peacekeepers, and international agencies would have contributed to even greater reductions in mortality.

Survey Limitations

As with all surveys, our study has limitations. Several of these relate to the problems of applying cluster sampling methodology, a tool originally designed for vaccination coverage surveys, to the collection of mortality data. These limitations have been described well elsewhere. 11,24,25

We relied on MOH population data for the random selection of health zones. Wherever possible, we also used MOH data for the selection of individual clusters. When such data were not available, the relative size of smallest units was crudely weighted using estimates of local leaders. No independent confirmation of these estimates was conducted due to time constraints. It is unclear whether inaccuracies in MOH data or in the estimates of local leaders could have biased our results, but we have no reason to believe that these would have led to significant systematic errors.

Our choice of a method to select households was bound by time, logistical, and resource constraints. Systematic random sampling was our preferred method of household selection, but this was only feasible in 16% of the smallest sampling units because most were dispersed, rural communities. In such settings, we used the WHO/EPI random walk method, which has been used widely in other mortality surveys.Reference Spiegel and Salama26, Reference Depoortere, Checchi and Broillet27 Articles published since our fieldwork have suggested amendments and alternatives to this method, although each have their own limitations and biases.Reference Rose, Grais and Coulombier24, Reference Grais, Rose and Guthmann28

We used the current household census method for the ascertainment of events and the population at risk during the recall period. Although the past household census method is preferred, our own experience indicated that it significantly increased the time at each household, posing problems for staff in remote and insecure areas. The current household census remains a valid but not ideal method of household data collection.Reference Checchi and Roberts11, Reference Rose, Grais and Coulombier24

There was a high proportion of empty households in both strata. A limited assessment during the 2002 survey found that mortality was 43% higher in empty households than in households in which occupants were home at the time of the initial household visit.Reference Roberts and Zantop4 Logistical issues prevented a similar assessment during the most recent study; therefore, we do not know whether this pattern holds true in DRC today. We tried to reduce exclusions by surveying early morning or late afternoon, and avoided interviewing on public holidays and on Sundays in Christian communities. Ideally, all surveys should return to empty households at a later date and avoid replacement with a neighboring household. If this is impractical, then a subset of empty households should be revisited to gauge the likely direction and magnitude of this bias. It is a clear limitation of our survey that we did not use these measures.

Fever/malaria was reported to be the number one killer in DRC and our inclusion of 2 wet seasons in the recall period could overestimate mortality rates. This should not, however, influence comparisons between our 2004 and 2007 surveys because recall periods for both surveys covered exactly the same months. Cause of death data were as reported by interviewees and must be interpreted with caution. Deaths due to trauma (violent and accidental) and measles and deaths during the neonatal period are likely to be accurate, although the latter may be underreported. Our cause of death data should only be used in a general sense to alert planners to possible health priorities, and should be followed by more accurate local assessments to determine cause-specific mortality rates.

This is the first IRC survey in DRC in which zones were not omitted from the sampling frame in advance due to insecurity. Except for 1 small, militarized zone in the west, all of the health zones were eligible for selection. Nonetheless, as with every survey in conflict settings, our results could have underestimated mortality because of survival bias, underreporting of infant deaths, and exclusions of segments of the population from the survey; large areas in 1 zone in North Kivu became unsafe to visit after surveying began, and a military camp and a remote, rural village could not be accessed.