On the estimation of population cause-specific mortality fractions from in-hospital deaths

Well-functioning vital registration (VR) systems are essential to monitor health progress and inform health policy. They are a critical input for public health analyses, understanding epidemiological patterns, and allocating scarce resources for public health and medical care. Most developed countries have complete or near complete VR systems, with reasonably accurate cause of death data. However, VR systems in low- and middle-income countries (LMICs) are generally incomplete in reporting causes of death, and the need for representative, quality data on leading causes of death is even greater [1].

Reports of hospital deaths are available from nearly all LMICs. In the absence of a comprehensive VR system, many countries use hospital deaths as the basis for their cause of death (COD) reporting and the identification of health priorities [2]. The advantage of hospital-based reporting of deaths is that hospital practitioners have the investigative capacity to discriminate between conditions with common sets of symptoms, although this does not necessarily mean that causes of hospital deaths are being accurately diagnosed [3, 4]. The disadvantage of using hospital data to extrapolate cause-specific mortality fractions (CSMFs) to whole populations is that hospital deaths are a biased sample of all deaths and thus are unlikely to accurately represent the distribution of deaths by cause in the population.

Murray et al. show that this selection bias associated with age and sex can, however, be addressed once the probability of a person of a particular age and sex dying in hospital from a particular disease or injury is known [5]. It is then possible to calculate the population-level distribution of deaths by cause, age, and sex. Put simply:where H_asj is the number of deaths in hospital from age group a, sex s, and cause j; D_asj is the number of deaths in the population from age group a, sex s, and cause j; and P_asj is the proportion of all population deaths from age group a, sex s, and cause j which occur in hospital. We can therefore estimate D_asj by dividing H_asj by P_asj. Summation of D_asj over a and s gives the estimated total number of population deaths from cause j, from which the population CSMF, the primary quantity of interest for guiding public policy, can be directly calculated.

This approach was applied to deaths in Mexican hospitals between 1998 and 2005 to estimate population CSMFs for the population [5]. VR during this period was estimated to be 90% complete in Mexico and 97% of deaths were certified by a physician [6]. As a result, P_asj could be calculated precisely given the high levels of VR completeness, and CSMF extrapolations could be validated against VR, assuming it is of reasonable diagnostic accuracy. The authors found that even when reducing the amount of VR data to calculate P_asj to 9%, the relative error between the true and estimated CSMF was only 12%. The key unknown affecting the generalizability of this approach, however, is the stability of P_asj under variable population circumstances.

The major problem with this method, however, is that P_asj can be difficult to determine. For example, a study in Malawi found that estimating the proportion of child deaths that occur in health facilities was complicated by potential overreporting of child deaths in-facility due to social stigma [7]. To address this issue, Murray et al. suggest using P_asj values of neighboring countries. While this approach may be attractive due to its relative simplicity, its scientific basis is questionable. Murray et al. also argue that P_asj varies depending on regional socioeconomic status and other factors, such as education. In their Mexico dataset, they showed that the proportion of deaths in-hospital due to HIV/AIDS ranged from about 0.35 for the least literate to 0.7 for the most literate populations.

As a solution to the variability of P_asj between regions and the difficulty in determining it, we propose an empirical methodology (as opposed to indirect approach used in Mexico by Murray et al.) that allows researchers to use hospitals deaths to predict population CSMFs while correcting for factors that affect place of death. We applied this approach to two sites in the Philippines and Bangladesh to ascertain if the method might be a valid basis to derive national COD estimates. We tested this hypothesis by comparing CSMFs derived from the method with population CSMFs derived from verbal autopsy (VA) using the Tariff 2.0 method and national Global Burden of Diseases (GBD) estimates [8].

We identified 4288 and 3725 adult deaths in the Philippines and Bangladesh, respectively. The proportion of deaths said to have occurred in-hospital from the VA interview varied substantially (Table 1). More deaths in the Philippines occurred in-hospital than in Bangladesh (Table 1). In the Philippines, a majority of maternal deaths occurred in-hospital, and a large portion of deaths due to stroke, other NCDs, acute infections, chronic infections, and injuries also occurred in-hospital. In Bangladesh, a large portion of maternal and injury deaths occurred in-hospital.

The optimal model (best AIC) for each COD was approximately compliant with the 10 events per covariate guideline used to avoid overfitting. The covariates available for each cause varied depending on the number of deaths and model convergence. In both countries, age was consistently negatively associated with death in-hospital (Tables 2 and 3). In the Philippines, time to hospital less than 30 min was consistently the strongest predictor of death in-hospital. At least 8 years of education was also a strong predictor in the Philippines. Disability was a consistent predictor of death out-of-hospital in Bangladesh. Wealth, in terms of a wealth score or quintile derived from the list of available wealth-related covariates in Additional file 7 and PCA, was significantly associated with in-hospital death. However, given the logistical difficulties of deriving this value, wealth was excluded as a predictor in all other analyses. CSMFs using the wealth index are shown in Additional file 8.

The results of applying the various methods to estimate community CSMFs in the Philippines are shown in Table 4. Other non-communicable diseases (NCDs), excluding stroke, was consistently the leading COD across the different estimation approaches, followed by stroke and acute infections. The fraction of deaths attributable to cancer, diabetes, chronic infections, and maternal deaths varied across methods. Maternal deaths clearly had the lowest CSMF, as might be expected, accounting for less than 5% of deaths.

Community CSMFs, assigned by the Tariff 2.0 VA method, and hospital CSMFs, assigned by medical record review, were similar for injury, maternal, and other NCD deaths. The CSMFs estimated from the DPHD method that considered covariates related to the probability of death in hospital were greater than the hospital estimates for cancer and chronic infection deaths. Chronic infection from the DPHD method had a wide confidence interval, likely due to the low probability of hospitalization for chronic disease deaths. The DPHDNC method estimates, which did not consider covariates related to death in-hospital, were similar to the estimates from the method with covariates, with the only notable exception being that the DPHDNC method estimated more other NCD deaths.

The national GBD estimates showed closest similarity with the community estimates, with the exception of injuries, stroke, and other NCDs. The community estimated more injury and stroke deaths while GBD estimated more other NCD deaths. Based on our experience with implementing the Tariff 2.0 method in several countries, a potential explanation for the large difference between community and GBD estimates for other NCDs and stroke could be incorrect translation or misinterpretation of the word “paralysis” in the VA instrument, leading to overreporting of paralysis and therefore a disproportionate number of stroke deaths predicted by Tariff 2.0.

Comparison of CSMFs from application of the same methods to Bangladesh data is shown in Table 5. Other NCDs and stroke were the leading COD across all methods. Other causes displayed moderate variation across the methods. In particular, while community and hospital estimates were similar for acute and chronic infections, community CSMF estimates were greater than the hospital estimates for cancer, diabetes, and stroke deaths and less than the hospital estimates for injury, maternal, and other NCD deaths. The DPHD method estimates were generally between the community and hospital estimates. Again, the GBD estimates were closest to the community estimates, with community estimating more stroke deaths and GBD estimating more other NCD deaths, potentially for the reason as described above. The DPHDNC method estimates were similar to the estimates from the DPHD method, with the only notable exception being that the method without covariates estimated more injury deaths. Overall, estimates across all methods in Bangladesh displayed general agreement when considering the confidence intervals.

We have proposed a novel method to estimate cause-specific mortality patterns from in-hospital mortality statistics, which are generally available in most low- and middle-income countries. The method builds upon other approaches that adjust hospital mortality data for the population by taking an empiric, rather than indirect, approach and offering a simple application to hospital catchment areas. Application of this method identified stroke and other NCDs as the most frequent COD in the Philippines and Bangladesh. Hospital death certificates, as well as application of Tariff 2.0 to the hospital catchment area, also identified these two causes as the most frequent COD. The rank order of importance of the remaining causes of death varied across the methods under comparison, although the DPHD method estimates generally estimated a CSMF somewhere between that resulting from hospital data and the community. The CSMFs from the models with and without covariates were similar across all causes in both populations. Weighting by the probability of a particular cause occurring in-hospital potentially adjusted for the overrepresentation of causes that are more likely to occur in-hospital, such as maternal deaths, and the underrepresentation of causes that are more likely to occur at home, such as cancer deaths.

The likelihood of a person dying in-hospital is affected not only by age, sex, and cause, but also by demographic factors relating to the culture and health care systems of the deceased [5]. Therefore, simply using the probability of death in-hospital from neighboring countries is unlikely to produce accurate cause of death estimates. The variability in the proportion of deaths in-hospital between the Philippines and Bangladesh demonstrates the difficulty in using the values of neighboring countries. Consequently, we adjusted for covariates (Additional files 1 and 2) that influence place of death, making the results potentially more applicable across countries. Our findings suggest that distance from hospital, marital status, and wealth influence the place of death. We found that time to hospital was the greatest predictor of death occurring in-hospital in the Philippines, while presence or absence of a disability was the greatest predictor of death out-of-hospital in Bangladesh. Wealth was a strong predictor of death in hospital in Bangladesh, but we were unable to use wealth in the final analysis due to the difficulty in determining wealth from medical records. Models that include wealth as a score or a quintile, or a binary variable indicating TV or fridge ownership are shown in Additional files 7 and 8.

In addition to the DPHD method that adjusts for covariates, we also estimated CSMFs without adjusting for covariates by simply using the proportion of people who died in-hospital for each cause according to Tariff 2.0 or medical record review. We found the CSMFs from this approach were similar to the method using covariates. While weighting by the probability of dying in-hospital without using covariates is far simpler than collecting the appropriate covariates for each hospital death, it is not possible for other countries or regions to use this method without conducting VAs on the hospital catchment area. Thus, adopting this estimation approach negates the time and cost savings of applying this methodology elsewhere.

Unfortunately, it is not possible to validate the DPHD method as neither Bangladesh nor the Philippines have complete VR systems. However, our methodology follows the work of Murray et al., who demonstrated that even in areas in Mexico with reduced VR coverage, hospital deaths weighted by the proportion of deaths in-hospital produced more accurate population CSMFs than hospital deaths alone [5]. We might therefore infer that the DPHD method in the Philippines and Bangladesh produced more accurate cause of death estimates than the hospital CSMFs.

It is difficult to identify an appropriate set of validation data against which to test the predictive performance of our approach to estimating cause of death patterns. Probably the most appropriate comparator data set are the country-level GBD estimates since they make extensive use of all available mortality and covariate data, appropriately adjusted for known biases and modeled to describe the likely COD patterns in 195 countries [15]. Since the GBD study does not provide sub-national disease estimates for the Philippines or Bangladesh, we do not have a better validation dataset for the COD distribution in study sites. We found that the GBD showed general agreement with the DPHD method but showed closer agreement with the community results. We cannot determine the relative performance of the DPHD method versus community, but we can assert that the method can be far more easily applied than conducting VAs. As VR systems improve, the implied cause distribution in the population might be compared with that of the DPHD method as one external measure of plausibility.

While the DPHD method produces less-biased estimates than hospital deaths alone and can be more easily applied than VA, there are implementation issues that could affect the validity of estimates. First, we attempted to use demographic covariates, such as age, sex, marital status, and education, that are widely available in hospital medical records and would be influential in determining place of death, but some countries may have difficulty collecting this information in a systematic fashion or the covariates may be unavailable. Without these covariates, population CSMFs cannot be calculated. Second, a recent systematic review questioned the quality of hospital diagnoses, which for this method has assumed to be a “gold standard” [16]. Poor hospital diagnoses would bias the extrapolation to the hospital catchment area. Similarly, VA was used to assign individual causes to community deaths. VA is the only plausible way to assign causes to these deaths and has reasonable performance at the population level, but less so at the individual-cause level [8]. However, as discussed earlier, some of the differences between the community and GBD estimates for various causes might well be attributable to misinterpretation of the VA questionnaire by the respondents. Third, to consider the international variations in diagnostic practices, capacity, and skills, we condensed COD into eight broad causes, which limits the granularity of extrapolated CSMFs. Nonetheless, the broad cause categories that we have proposed still provide valuable information on the likely stage of the epidemiological transition in the population to which it is applied. Future research, where possible, should attempt to use more specific causes of death. In addition, future studies should consider the difference in diagnostic accuracy between hospital death certification and VA, which we did not take into account in this study. Fourth, while we could calculate confidence intervals for the modeled CSMFs in Bangladesh and the Philippines, this would not be possible in other countries without the variance-covariance matrix of the beta coefficients, which necessitates data from a population survey. However, model confidence intervals may be less of a limitation given recent controversy over the use of confidence intervals in global health research, given the fact that variability is likely to be at least as much affected by other factors related to data quality as model specification [17]. Finally, no true validation dataset exists in which to validate the DPHD method. We attempted to use the most available and robust data to assess its performance. Given the limitations of the data, we were unable to test whether the method could be used within or between countries. The Philippines and Bangladesh have relatively similar mortality profiles. Future research should attempt to validate this method against data sources for other countries with somewhat different mortality profiles. Beyond the scope of this paper, there are potentially other data that could be used to test the performance of the proposed method, such as patterns of age-specific death rates as well as application of the method to other countries without complete VR systems but with readily available hospital and census data, such as Brazil and Ecuador [18, 19].

Informed health policy requires reliable estimates of the leading causes of death for the entire population, not just for those who die in hospital given that the probability of dying in-hospital is likely to be age, cause, sex, and culturally dependent. Hospital CSMFs need to be weighted by the region and cause-specific probability of a death occurring in-hospital. It is possible to estimate these probabilities and use cause-specific logistic regression models to estimate the cause-specific probability of a death occurring in-hospital. These regression models could potentially extrapolate hospital CSMFs to the population in regions that are deemed to be culturally similar to the Philippines or Bangladesh, the two countries on which the method has been developed. For regions and countries that have not employed routine VA, and where deaths in-hospital are certified with reasonable accuracy, information collected from routine hospital medical records and death certificates can be used to generate, at no, or very low, cost, reasonably plausible population CSMFs.