This study developed and applied a standardized computer-based algorithm to assign child causes of deaths using nationally representative verbal autopsy (VA) data in Bangladesh. The results provide distinctive insights into patterns and trends of child causes of death at the national level for one decade. The empirical trends are of particular interest as child causes of death are often modeled for most LMICs [1, 2].
Our results corroborate the previous finding that pneumonia remains the top ranking cause of death among children below 5 years of age , despite the fact that the mortality rates of pneumonia had been declining significantly in Bangladesh. The proportional contribution of neonatal causes increased steadily during the study period. The proportion of under-5 deaths due to birth asphyxia/birth injury and prematurity/LBW rose progressively and significantly. The increase in their CSMR was also noticeable, although only the increase in the CMSR of birth asphyxia/birth injury reached statistical significance, possibly due to the competing significant reduction in the U5MR . Likely because of a similar reason, the cause-specific fraction and the CSMR of neonatal tetanus both decreased, although only the reduction in CSMR was statistically significant. The composition of the top three ranking causes transitioned from including none neonatal cause to including two neonatal causes, which signifies again the increasing relative importance of neonatal mortality as under-5 mortality rate continues to decrease .
The CSMR of all postneonatal causes dropped significantly between 1993-1994 and 2004 except for other possible serious infections. Some of these changes, such as those in measles and diarrhea, may be partly explained by the increase in the coverage of measles vaccine and oral rehydration salts [7, 11, 12]. Other changes, however, have a less obvious association with changes in intervention coverage. For example, despite the significant decrease in the mortality rates of postneonatal pneumonia, the case management of pneumonia did not seem to improve [7, 11, 12], although access to high-quality, low-cost antibiotics is suggested to have been increasing during this period.
Several approaches are in widespread use to estimate the variability of population-level estimates for complex multistage sample survey data. The Jackknife and Balanced Repeated Replication (BRR) methods need multiple primary sampling units per sampling stratum . These methods are both based on resampling of sample survey data. In this study, where each DHS sample had many levels of stratification, there were multiple strata where only one sampling unit was informative concerning the causes of death for children under 5. Resampling can accommodate stratification and clustering in multistage sampling data, and a more general approach relative to the Jackknife and BRR was used in our case, where an additional complexity was introduced due to the cause reallocation from the comorbidity of pneumonia and diarrhea to each individual cause as a nonlinear function of population estimates.
We developed a standardized algorithm to assign child causes of deaths using signs and symptoms collected through the VA studies to ensure comparability of the trends. The original algorithms applied in each study were systematically reviewed and key differences in cause categorization, hierarchy, case definition, and amount of data collected were compared to inform the development of the standardized algorithm. The standardized algorithm, primarily based on the 2004 cause categories and hierarchy, guarantees comparability of the trends by only including information commonly available across all studies. However, this was realized at the expense of losing additional information collected in the later studies.
When determining the cause categories, instead of mapping the causes to specific codes based on the ICD-10, we employed the CHERG cause categorization, which is in principle consistent with the ICD-10 rules. However, the CHERG classification does emphasize the relative public health importance of major child causes and encourages linkage with the relevant health interventions .
When presenting the final causes, we chose to combine possible pneumonia and possible diarrhea with pneumonia and diarrhea, respectively, for several reasons. First, based on the case definitions, we considered that possible infections were likely to be true cases of pneumonia/diarrhea with less severe symptoms. Second, after combining the possible diagnosis with the confirmed causes, the time trends were more stable. This practice is also consistent with previous approaches taken by CHERG [1, 2]. However, the resulting trends in pneumonia need to be interpreted with caution. Further investigation shows that possible pneumonia is responsible for 12% to 26% of the deaths coded as pneumonia deaths among children aged 1 to 59 months old but 59% to 71% of the deaths coded as pneumonia among neonates. We suspect that the high proportion of neonatal possible pneumonia could be due to the inclusion of other serious infections, such as neonatal sepsis. As a common caveat of VA studies , additional misclassification errors may also exist in our results.
Comorbidity between diarrhea and pneumonia were reallocated among the two causes based on their relative importance in this study. Alternatively, cases with this type of comorbidity can be treated solely as pneumonia deaths. We took the former approach based on biological and medical considerations. However, more research is clearly needed to further determine which method is more appropriate. In fact, CHERG has an ongoing activity to examine the comorbidity patterns between pneumonia and diarrhea, which may contribute some knowledge to this area in the near future.
We estimated that malnutrition was responsible for 6% to 8% of under-5 deaths in Bangladesh in the study period. However, malnutrition could contribute to additional under-5 deaths as a risk factor . In our sensitivity analysis where multiple causes were initially allowed, we found that comorbidity between malnutrition and any of the four infectious causes (diarrhea, ARI, measles, and other serious infections) exist among 42% to 55% of these deaths. Therefore, the estimated contribution of malnutrition to under-5 deaths needs to be interpreted carefully.
Several limitations are acknowledged regarding the present study. First, despite our attempts, several important causes, such as meningitis and neonatal sepsis, were not classified due to a lack of necessary symptom data in the two earlier studies. Moreover, unspecified causes contributed to 18% to 25% of under-5 deaths in Bangladesh between 1993 and 2004. It is suggested that more symptom data, especially relating to serious infections, should be routinely collected in future VA studies to facilitate ascertainment of additional causes.
The case definitions used herein were validated in Bangladesh, Nicaragua, and Uganda, and their validity has been shown to be reasonably good [10, 19]. Similar hierarchies, though not validated, have been applied to the three Bangladesh VA studies originally and to a VA study conducted in parallel with the 2006 Nepal DHS . Our standardized hierarchy, however, has the inherent limitations of any hierarchical process. In particular, the resulting cause-specific fractions are quite sensitive to the tier in which the causes were assigned [21, 22]. For example, for comparison purposes, we assigned prematurity/LBW after other possible serious infections. The prematurity fractions were reduced by about 30% across all three studies. However, the hierarchy doesn't affect all deaths in the studies. In fact, our sensitivity analysis reveals that more than half (52%-54%) of the deaths were either due to a single cause or considered as belonging to "unspecified causes". Their cause fractions will remain the same irrespective of the application of hierarchies. Among the other half of the deaths with multiple diagnoses, many deaths could have been assigned to the same causes during medical certification.
Further validation research, such as the ongoing Grand Challenges in Global Health Initiative #13 Study done by the Population Health Metrics Research Consortium (PHMRC), may help determine a better computer-based algorithm. However, preliminary results of the PHMRC show that infectious causes, such as pneumonia, are still among the few causes that are hard to assign even after applying the advanced machine-learning theory. Whether these validated algorithms can be readily applied to secondary data is questionable since only limited information on signs and symptoms are available. In addition, the external validity of these algorithms in other countries and settings still needs to be shown.
Limited by the validity of the algorithm, the absolute level of the estimated cause-specific mortality rates may not be accurate. In other words, our uncertainty ranges did not take into account the unknown uncertainty in the cause of death assigning process. But with the employment of the standardized algorithm over the three datasets, the trends in the cause-specific mortality rates should be reasonably reliable. The uncertainties of the trends have been quantified by incorporating known uncertainties in the complex survey design and the U5MR. The values of these trends are further appreciated considering the fact that they originated from nationally representative empirical data. In the absence of a more reliable alternative, such information should start to be utilized to facilitate child health policymaking and resource allocation.
In the past, few empirical data on child causes of deaths have been available in LMICs, but recently more data are being collected and shared. Bangladesh is among a number of countries that have had at least one DHS with the VA module, and more countries are either planning on or considering including the VA component in their future DHSs. In addition, cause of death data collected through other sources, such as the International Network for the Demographic Evaluation of Populations and Their Health in Developing Countries (INDEPTH) and the Mozambique post-census survey applying the Sample Vital Registration using Verbal Autopsy (SAVVY) methodology, could all contribute to a better empirical understanding of cause of death . Often, however, the heterogeneity in cause of death estimation could also be due to differences in methods for assigning or ascertaining causes [2, 23]. Standardized algorithms are becoming an indispensible tool to generate child causes of death estimates that are comparable across time, countries, and settings . Computer-based algorithms have the advantage of being objective, feasible, and affordable, and may better serve the needs of LMICs as compared to physician review. The current exercise is among the efforts to develop such standardized tools with the hope that more research of this type would be stimulated.
The resulting trends in child causes of death also provide a platform to link with trends in intervention coverage in Bangladesh. Close examination of these linkages, using packages like the Lives Saved Tool (LiST) , would help us better understand what drives the changes in cause-specific mortality and to identify mechanisms that work in the context of Bangladesh to reduce child mortality. These successful experiences or lessons learned can be shared with other countries to help accelerate their progress toward Millennium Development Goal 4.