Estimated probabilities of having a GC as the cause of death decreased 40.4% from 1996 to 2016 in Brazil and the decrease was particularly important for the level 1 codes (58.4% over the period). The sharp decline presented by this group may be related to the investigation in hospitals and at home of an important part of GCs from this group, the R-codes, which has been performed by the Ministry of Health (MoH) since 2005 [5,14]. States classified in low or medium SDI tertiles were responsible for the most important decline. Greater improvement in Brazil´s less developed states might be related to investments in the public health care system [15]. Although the relationship between SDI and rates was maintained in 2006-2016, rates of level 1-2 GCs in the three different state groups converged over time, indicating reduced inequalities in data quality due to major GCs. This reduction may have been related to the intervention of the MoH focused on poorer areas, and therefore leading to a reduction in those codes [6].
The decrease of major GCs occurred in almost all Brazilian states, and this encouraging result also indicates progress in strengthening cause-of-death statistics in the country over the past 20 years. Our findings are supported by the results of the classification system developed for the GBD2016 study - the GBD data quality rating. The calculation of this rating is based on the following variables: completeness of death registration, proportions of deaths not coded to GCs levels 1 or 2, and fraction of deaths assigned to detailed GBD causes. The score for Brazil improved from 68.9 in 1995-1999 to 82.5 in 2010-2017, although it remained as a 4-star country in this 1 to 5 star classification. In the latter period, some high-income countries like Germany (score=84.0), the Netherlands (83.3) and Japan (81.3) were also classified with 4 stars. On the other hand, in South America, Chile and Colombia were classified with 5 stars, but Uruguay, Paraguay and Argentina were also classified with 4 stars [10].
Although with a decreasing trend, GCs still represent an important percentage of total deaths in the country, influencing the quality of mortality information. Despite the improvement in cause of death quality, R-codes and other ill-defined causes such as heart failure or septicemia, which contain no information of the underlying cause, remain frequent causes. High proportions of deaths due to these causes indicate not only worse quality of cause of death statistics, but also lack of medical care, especially among the economically disadvantaged population [16]. Interesting, proportions of pneumonia unspecified increased in 2016 for men and women, probably related to the observed increase over time in age-specific mortality rates from specific infections of the lower respiratory tract in the population aged 70 years or over [17].
The highest proportion of GCs occurred among elderly people, especially those over 80 years (49.6%), and were due mainly to level 4 GCs. Higher proportions of ill-defined causes among the elderly are probably due to the higher number of comorbidities, such as neoplasms, hypertension, diabetes and other cardiovascular diseases, making it difficult to provide information on the underlying cause of death during the completion of the DC [18, 19, 20], even after the use of verbal autopsies [21]. On the other hand, children aged 1-14 years also had a high fraction of major GCs, more than 20%. High proportions of GCs in this age group, although in accordance with the results of Naghavi et al.(2010) [8] in other settings, are of concern as it is very important for prevention to correctly identify the leading causes of premature deaths.
Regarding the place of death, GCs were registered as the underlying cause of death in almost 50% of home deaths and in more than 38% of those which occurred in a hospital. Although previous studies in some Brazilian cities showed that ill-defined R codes were negatively associated with hospital deaths [22, 23], the percentage found for in-hospital deaths in this study can be considered relatively high. In hospital settings diagnostic procedures are more available, and we expected that better defined CODs would be declared much more frequently.
The highest proportion of GCs in home deaths may be due to difficulties of diagnosing or a lack of knowledge of the diagnosis by the physician. This can be explained by the fact that the certification of a home death is more difficult than deaths occurring inside hospitals due to the difficulty of defining a diagnosis for reasons already mentioned. The majority of home deaths in Brazil had medical assistance before death [24]. This finding reinforces the importance of continuing the investigation of household deaths until a culture valorizing the importance of the correct DC information is establishing among physicians.
Proportions of deaths coded with GCs were similar among categories of certifying physicians, although we had originally expected a smaller proportion to be present when attested by the attending physician, and by those from SVO and forensic institutes. But for major GCs, the proportion of GC level 2 in deaths certified by forensic institute physicians was higher than that found for other physician categories. These findings probably result from deaths from injuries whose intent remained undetermined despite being certified by a coroner.
In order to investigate the impact of socioeconomic conditions on rates of different types of GCs over time, we have presented three sets of calculations at the state level. First, we presented age-standardized rates by SDI in two periods, 1996-2005 and 2006-2016. These results indicate that while there has been a decline of GC levels 1-2 in both periods, mortality rates coded to GCs levels 3-4 increased as SDI increased across states and time during 1996-2005 but decreased as SDI increased during 2006-2016. Second, we presented linear regression models in two separate periods for analyzing coefficients of SDIs and completeness related to GCs. The inclusion of completeness was intended to address whether this variable impacted on GC occurrence, as there is evidence of an association of improvements in completeness with declining percentages of deaths being reported as ill-defined [4]. Finally, extended temporal analysis of states classified in tertiles of development according to state of residence ranked by estimated SDIs in 1996 indicated substantial differences in performance between low and high SDI tertile states in 1996-2005, with the gap closing substantially in the 2006-2016 period.
Our results of an inverse association between levels of SDI and rates of levels 1-2 GCs in both periods are consistent with recent publications focused on ill-defined causes of death. Higher proportions of R-codes from chapter 18-ICD10 are more frequent among people who had a lower educational level or less access to health services [25, 26, 27]. On the other hand, the unexpected significant direct relationship between SDI and mortality rates due to GC levels 3-4 in 1996-2005 constitutes the most striking finding from our study. To our knowledge this finding was not reported before in the literature. We also observed a significant direct relationship between completeness and proportions of GC levels 3-4 in the first period, i.e., these GCs were higher when completeness increased.
The GBD 2017 study aggregates locations into 5 SDI groupings generated by applying quintile cutoffs from the distribution of national-level SDI for countries with populations greater than 1 million in 2017 to estimates of SDI for all GBD locations. Brazilian states were classified into 3 of those groupings: Low-middle SDI, Middle SDI and High-middle SDI [10]. We adapted this procedure by applying tertiles cutoffs to estimates of all state-level SDI in Brazil in 1996 to better contextualize the Brazilian reality. When analyzing temporal changes of GC rates across those tertiles of SDI, we observed that the increase of rates of GC levels 3-4 were concentrated in the low SDI tertile states in 1996-2005, and this slight increase occurred simultaneously with the increase in completeness in this group of states, and greater decrease of levels 1-2. It should be due to expanded primary care, with better knowledge of patient problems prior to the admission that led to death [27].
We suppose that states with better completeness and socioeconomic standards have more hospitals, and physicians were able to fill out the DC with GCs levels 3-4 such as pneumonia, diabetes or stroke unspecified, instead of an R code from level 1. The growing prevalence in poorer areas of some of these GCs such as diabetes of unspecified type, with higher burden in the North and Northeast regions [28], could also had contributed. Another hypothesis is related to the correction of undercount of deaths in this study, as this adjustment was based on the assumption of a similar COD distribution in deaths recorded and not recorded in SIM, which is probably not correct for some causes. Different distribution of causes among registered and missing deaths were observed in Minas Gerais State, Brazil, using the verbal autopsy technique [29]. Further detailed investigation of this aspect can help us to identify specific issues of this subject.
Overall, the predominance of level 4 GCs in the country in recent years is encouraging, as these causes have less impact in public health. According to Anaconda, a tool for checking the quality of mortality data, GC levels 1-3 are classified as higher severity unusable causes, given their higher risk of misguiding public policies to avoid premature deaths. Level 4 codes are not included in the broad category of unusable causes by this classification. Furthermore, level 4 GCs include codes that may require access to greater diagnostic sophistication and considerable equipment to accurately determine the COD [30].
This study has two major limitations. First, GC mortality is corrected for under-registration using the estimates of all-cause mortality from the GBD study, which can be imprecise [10]. Furthermore, as was previously stated, the correction for cause-specific rates was based on the assumption of a similar COD distribution in deaths recorded and not recorded in the SIM, which may not be the case [29]. Secondly, to analyse socioeconomic differences at the regional level, we used the sociodemographic index (SDI), a composite indicator proposed in more recent GBD studies as a measure of socioeconomic deprivation. We chose the SDI over other deprivation measures as it is frequently used in various studies as an indicator of the socioeconomic level of countries and subnational divisions, and uses similar component indicators for countries and states [10], which facilitates comparisons in the global context. Nevertheless, there are no set criteria for judging the quality of estimates of measures of deprivation, and hence it is difficult to assess whether the SDI is the best indicator. However, it summarizes three important dimensions of human development and is calculated in years from 1990 to 2017, according to methodological adjustments made for the GBD 2017 study [10].
The analysis plan developed here does not contemplate the potentially important socioeconomic differences by sex in completeness and trends concerning the quality of cause-of-death registry in Brazil, which therefore may also be considered as a limitation in this study. Our objective, however, was only to visualize the socioeconomic differences without control for confounding factors such as gender, access to health services and others. Additionally, if we were to stratify the analysis by sex, we would end up producing a considerably more detailed analysis, longer and more complex, and thus turning more difficult the task of understanding the principal objective, the description of the burden of GCs in Brazil.
In summary, this study provides a more accurate understanding of trends and magnitude of mortality coded to garbage codes according to levels and the ranking of specific groups. This information is of importance for health policy in Brazil as the proportions found in 2016 were still large enough to impact regional inequalities in well-defined causes of death without a redistribution of GCs. Although the GBD study redistribute those GCs among specific causes, this redistribution is based on statistical methods and has some limitations [10, 31]. As better correction is provided by empirical data [32], this study might also contribute to the recent effort made by the Ministry of Health who in 2017 implemented important interventions in 60 municipalities throughout the country, as part of the Bloomberg Data for Health Initiative, in which hospital deaths with garbage codes were investigated extracting information from hospital records using standardized questionnaires, which then were reviewed by physicians in order to identify the underlying or specific causes. This initiative, reinforced by adequate physician training on the filling in of the DC, has significantly improved data from the SIM [33]. So, it is imperative to have a comprehensive knowledge of GC distribution in the country, therefore permitting further advances in this intervention and the subsequent empirical corrections.