Introduction
Cause-specific mortality data are of paramount importance to describe the health profile of a population, to set priorities for health policy makers, and can be used to evaluate the impact of preventive interventions. Moreover, cause-specific mortality represents a commonly adopted end-point for epidemiologic studies and clinical trials, and constitutes an essential information source to build disease registries.
Usually, reported mortality data are limited to a single cause of death, the underlying cause of death (UCOD), which is selected from all diseases mentioned in the death certificate according to international coding rules. An additional approach is represented by the analysis of all conditions reported in the certificate (multiple causes of death – MCOD), to assess any mention of a disease irrespective of its selection as the UCOD.
The adoption of MCOD analyses is necessary to capture the multiple diseases that can lead to chronic disease mortality, such as diabetes-related mortality [
1,
2]. Diabetes is associated with several conditions resulting in an increased risk of death, mainly cardiovascular disorders, renal failure, and infectious diseases; furthermore, recently the association has emerged between diabetes and mortality from various forms of cancer [
3,
4], and from chronic liver disease (CLD) [
5].
Important issues associated to certification practices arise when dealing with diabetes-related mortality. The first is the mention of diabetes on the death certificate: although diabetes might not be reported because it really did not contribute to death, previous studies found that patients’ demographics, place of death, role of the certifying physician, duration and treatment of diabetes, and associated diseases influenced the probability of mention of diabetes [
6‐
11]. However, such studies were limited to the US and UK, and were mainly carried out on small samples of known diabetic subjects. Another problem is placement of diabetes within the death certificates: it heavily influences the probability of its selection as the UCOD, and has been demonstrated to largely vary by country [
12]. MCOD data allow problems to be identified in the process of recording and elaborating information on death certificates [
13], and are fundamental to elucidating heterogeneity between countries in the selection of chronic diseases, including diabetes, as the UCOD [
14,
15]. Indeed, mortality rates from diabetes, judged only from the UCOD, can fluctuate over time due to trends in certification practices; since MCOD provide the most consistent data on time trends [
16], both MCOD and UCOD mortality rates should be examined to properly interpret changes in cause-specific mortality [
17,
18].
Lastly, MCOD can be analyzed to explore associations between different diseases leading to death; however many intricacies arise when associations found between conditions mentioned in the certificate are assumed as causal relationships at the population level [
19].
The present study explores the limits and potential of MCOD analysis in the assessment of the mortality burden from diabetes in northeastern Italy, with the four following specific aims:
1.
to examine patterns of the mention of diabetes in different sections of the death certificate, and the associated probability of selection of diabetes as the UCOD;
2.
to compute mortality rates and proportional mortality for diabetes, comparing figures based on the UCOD and on MCOD;
3.
to measure the rate and test factors associated to the mention of diabetes in a subset of death certificates of decedents from a cohort of known diabetic subjects;
4.
to assess the potential of MCOD data in the analysis of associations between different diseases reported in the certificate, even when the knowledge about such relationship is still limited within the medical community. As an example, results on the association between diabetes and CLD in the whole regional mortality archive were compared with those of a study on mortality from CLD in a cohort of diabetic subjects [
5].
Discussion
The study provides a comprehensive picture of the use of MCOD in investigating the mortality burden related to a chronic disease. The first issue regards diabetes reporting in death certificates of decedents affected by the disease. Physicians may not record diabetes on the death certificate because they could be unaware of the disease in the patient, may not consider that diabetes contributed to the patient’s death, or may not have listed diabetes because of space constraints [
7]. Physicians completing death certificates mention only conditions considered to be instrumental in causing death, not all prevalent diseases at death [
25]. According to a review, the median proportion of diabetes reporting in any position of the death certificate among decedents with known diabetes was only 43 % [
26]. The recording of diabetes in US death certificates did not increase from 1986 to 1993, being less than 40 % among decedents with a history of diabetes [
6]. In the TRIAD study, 39 % of diabetic subjects had the disease recorded; diabetes was less frequently reported for all causes of death other than cardiac diseases, especially cancer [
7]. In the US, the mention of diabetes was higher if the certifying physician was the primary care physician [
8], or if the place of death was the patient’s home [
9]. In studies from the UK, diabetes was mentioned in 42-43 % of death certificates of known diabetics, being associated with an increased duration of diabetes and insulin treatment, increasing age, female gender, low social class, and a cardiovascular underlying cause of death [
10,
11]. The present findings show a higher proportion of diabetes reporting (52 %) with respect to studies from the US or the UK. A problem of over-designation of diabetes as a cause of death exists if the certifying physician lists all diseases affecting the decedent, without considering their real role in causing the death. On the other hand, the present data show that diabetes was not mentioned in a substantial proportion of certificates where cardiovascular disorders such as ischemic heart diseases were selected as the UCOD. Furthermore, the study confirms that many variables are associated with the mention of diabetes: patient’s gender, long duration of disease, circumstances of death (for subjects dying at home certificates are generally filled by the family doctor or by a community health physician), and co-existing diseases leading to death. Unfortunately, we did not have data to test the role of diabetes treatment on the mention of the disease (no treatment, only oral anti-diabetic drugs, insulin). Moreover, we had no measure of the specificity of diabetes reporting; according to the few studies available in the literature, specificity was 98 % in the Rancho Bernardo study [
9], and the positive predictive value was 99 % in a sample of death certificates in France [
27].
The second major issue is the quality of cause-of-death statements in certificates with mention of diabetes. It is often difficult for the certifying physicians to decide whether to report diabetes in Part I of the certificate, which indicates that diabetes directly caused death, or in Part II, which suggests that diabetes contributed to death but was not part of the sequence of events directly leading to death [
20]. As an example, Taiwanese physicians were much more likely to report diabetes in Part I (70 %) than their counterparts in Sweden (21 %) and in the US (36 %) [
12]. When diabetes is recorded in Part I, there are two major types of improperly filled cause-of-death statements: reporting more diagnoses per line, and incorrect causal sequence among the reported diagnoses, usually resulting in an UCOD selected by the ACME software different from the underlying cause chose by the certifying physician. In previous studies, about three-quarters of the incorrect causal sequences involved incorrectly reporting other conditions as the cause of diabetes, mainly hypertension and acute myocardial infarction. A less frequent anomaly was diabetes incorrectly reported as the cause of other diseases [
15,
28]. In our database, diabetes was most frequently mentioned in Part II of the certificate. When diabetes was reported on the line reserved for the underlying cause in Part I as the only or the first reported condition, it was selected as the UCOD in most cases. Based on ACME decision tables, many diseases can be the consequence of diabetes, and rejected causal sequences starting with diabetes as the underlying cause were rare. The other types of error were much more frequent: reporting multiple diseases per line, and diabetes mentioned in other lines of Part I as due to other diseases, mainly selected circulatory diseases. In these latter cases, which can be regarded as a major flaw of death certification practices, diabetes was frequently selected as the UCOD. A main issue is the lack of training of the certifying physician. The development of electronic certification has been proposed to facilitate the process with online explanations, and to limit errors when completing the death certificate [
29]. However, due to the ageing population (the median age at death in the Veneto region was 78 years in males and 85 in females) and the associated increase in the incidence of multiple comorbid conditions, there may be no simple etiologic chain leading to the identification of a single underlying cause; instead, death often results from a complex interaction between multiple factors [
17,
29].
Within this context, official mortality data for chronic diseases should be provided based both on the UCOD and on MCOD. Age-specific mortality rates based on the UCOD were similar to those reported in the literature, but among older age classes, rates based on MCOD tended to be higher than previous reports. The ratio between diabetes reported as MCOD/UCOD displays heterogeneity by country: 4.2 in our study, 2.6 in France, 4.2 in UK, 4.5 in Sweden, and 3.1 in the US [
30]. Overall, diabetes was mentioned as MCOD in a larger proportion of overall deaths (12.3 %) than reported in other countries: 5.3 % in France in 2002 [
30], 5.1 % in England in 1995–2010 [
31], 10.6 % in Canada in 2004–2008 [
32], and 9 %, 10 %, 9 % in 2000–2001 in Sweden, Taiwan, and the US, respectively [
12]. This finding might be due to more recent data and an older population analyzed in the present study, and to a higher propensity to report diabetes in death certificates of elderly subjects.
The last issue is the association between different diseases mentioned in the death certificate. Different measures of association have been proposed in the literature [
33]. The simplest analysis would involve examining the frequency with which two conditions are reported together; estimation of Odds Ratios (OR) stratified by confounding factors such as age is usually necessary [
19]. We compared results on the association between diabetes and CLD in the regional mortality archive (Table
5) with those from the cohort of subjects exempt from medical charge [
5]. In this cohort, mortality from all CLD was higher than in the general population in both analyses restricted to the UCOD (SMR = 2.55), and in MCOD analysis (SMR = 2.55). Results remained unchanged if analyses were restricted to diabetics with long disease duration, confirming a role of diabetes in increasing the mortality from CLD. In MCOD analyses, mortality was higher for NVNA-CLD (SMR = 2.86) than for alcohol- (SMR = 2.25) or virus-related CLD (SMR = 2.17) [
5], a finding similar to Model 2 and 3 of the present results.
Redelings and colleagues have already reviewed possible bias affecting ORs estimated from mortality data [
19]. Any exposure that increases the likelihood that someone will die also increases the likelihood of inclusion in the study, generating a type of selection bias termed Berkson’s bias. Furthermore, chronic conditions are rarely reported in death certificates without multiple associated diseases, increasing the likelihood of spurious associations between them. Selection bias can be limited by choosing controls who died from diseases unrelated to the exposure [
19]. However, since diabetes increases the mortality risk from many causes, our first choice was not to restrict the selection of controls (Model 1); an alternative strategy was to select deaths from respiratory diseases (Model 2). Another strong bias possibly acting in our database was reporting bias due to common opinions among physicians completing death certificates, based on the biological plausibility of a connection between diseases [
19]. In our study, the mention of diabetes was increased in the presence of renal failure and selected circulatory diseases, and decreased by the presence of cancer (including liver cancer, often arising from CLD). This latter observation could be extended to many chronic diseases, since the overall prevalence of reported comorbid conditions is usually lower in deaths with an underlying neoplastic cause [
34,
35]. A further analytic strategy was therefore to stratify by major factors associated with diabetes reporting (Model 3); however, this could have led to overmatching bias [
19]. Furthermore, we had no data on completeness of the mention of CLD in mortality records. In view of all the above, caution is needed in examining associations observed between diseases in MCOD data, and multiple analytic strategies, such those adopted in the present study, should be explored to confirm results. The association between diseases reported in the death certificate depends on a complex interplay between different factors: the real etiological relationship between the diseases, current medical knowledge and beliefs, and certification practices. Such analyses could be useful to generate etiologic hypotheses to be assessed by other study designs. In our experience, preliminary findings from the MCOD archive showing the association between diabetes and CLD led us to perform the cohort mortality study.
The present study has both strengths and limitations in exploring the potential of MCOD analyses. Among the strengths, the study was carried out on mortality data coded according to ICD-10, with selection of the UCOD carried out by the ACME software. To our knowledge, this is the largest study investigating rate and factors associated to diabetes reporting in death certificates. Moreover, findings on the association between diseases mentioned in the certificate were compared with results of a cohort study. Among the limitations, we did not have a direct measure of the negative predictive value and the specificity of reporting diabetes in death certificates. Lastly, as in other countries [
1], over half of all diabetes deaths reported the code E149 (unspecified diabetes mellitus without mention of complications); therefore, examination of the breakdown of specific diabetic codes was not performed. It is worth noting that the selected UCOD is also often “diabetes without complication” when diseases such as renal failure are mentioned in the certificate. In these circumstances analytic strategies based on MCOD are essential to capture the real burden of mortality from diabetic renal disease [
36].
As a final remark, a paper published more than fifty years ago stated that although mortality statistics were not intended to give a comprehensive picture of disease prevalence among decedents, data based on MCOD were more informative than usual UCOD tabulations, especially for chronic conditions such as diabetes. Furthermore, some diseases, including liver cirrhosis, were found to be reported in association with diabetes more frequently than in overall deaths [
37]. In spite of these old observations, such analyses are applied in few countries, and the scientific literature on the potential and intricacies of MCOD remains limited.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
UF, GZ, GM: study design, interpretation of data, drafting of the manuscript. FA: statistical analysis, interpretation of data. CAG, MS: interpretation of data, revision of the manuscript draft. All authors read and approved the final manuscript.