Improving the Comparability of Diabetes Mortality Statistics in the U.S. and Mexico Free

Data for causes of death were from the National Center for Health Statistics National Vital Statistics System in the U.S. and from Secretaría de Salud in Mexico. Both sources maintain records for all deaths in the country, including multiple causes of death and standard sociodemographic characteristics. All deaths coded to “Diabetes Mellitus” in the MCD dataset between 1999 and 2001 in the U.S. (640,543 deaths) and between 2004 and 2005 in Mexico (182,796 deaths) were included in the analysis to maximize predictive power. Analysis was limited to deaths of individuals aged >20 years because very few diabetes deaths occur at younger ages. Deaths for which injuries were the UCD were excluded from the analysis because they were assumed to be correctly assigned.

We used multinomial logistic regression to estimate the relative risk ratios (RRRs) of a death, for which diabetes was recorded as one of the MCD, being assigned to diabetes versus cardiovascular diseases, other noncommunicable diseases, or communicable diseases as the UCD. The multinomial logistic regression estimates the RRRs for observing a dependent variable with more than two outcome categories as a function of independent covariates; the statistical basis and applications to epidemiological data are described elsewhere (15,16). RRRs were estimated for assignment to each of the disease clusters relative to assignment to diabetes. The independent variables in the regression were exogenous individual- and community-level variables listed in Table 1.

We used the coefficients of the multinomial logistic regression to predict the probabilities that each death would be assigned to each of these four disease clusters if the individual and community characteristics were those of a fixed reference category, as defined in Table 1. Although any fixed category could be used for the purpose of comparability, the reference category in Table 1 is expected to represent the highest access to specialized health care and, hence, the greatest likelihood of a valid cause-of-death assignment, above and beyond comparability. We controlled for age to avoid the confounding effects of factors that may also be associated with age. However, we did not use a reference age category in the adjusted (predicted) probabilities because some causes of death are expected to have true age patterns, which should be preserved. For the same reason, we adjusted the regressions for BMI and systolic blood pressure (SBP) but did not standardize these variables in the prediction stage to retain true epidemiological differences. Predicted probabilities were aggregated by sex, age-groups, and state to produce comparable estimates of diabetes deaths, and those of other disease clusters for each state. All analyses were conducted separately for the U.S. and Mexico to account for potential differences in the case fatality of diabetic patients in the two countries.

Excluding deaths from injuries, diabetes and cardiovascular diseases, respectively, comprised 33 and 39% of the UCDs among the universe of deaths with diabetes as one of the MCD in the U.S. and 67 and 14% in Mexico (Fig. 1). Cancers (10% in the U.S. and 6% in Mexico) and a number of other noncommunicable diseases accounted for the remaining proportion, except for a small proportion of deaths assigned to communicable diseases (4% in the U.S. and 3% in Mexico). Nearly 3 and 14% of all deaths in the U.S. and Mexico, respectively, are currently assigned to diabetes; 38 and 23% are assigned to cardiovascular diseases.

In both countries there were large discrepancies at the state level between the rates of diabetes as the UCD and the rates of deaths for which diabetes appears as one of the MCD on the death certificate (Fig. 2). For example, in 2001, age-standardized (using 2000 U.S. population aged ≥20 years) diabetes MCD rates in Hawaii, Utah, and Louisiana were between 10.0 and 10.7 per 10,000 but age-standardized diabetes UCD rates varied by a factor of >3 (1.8 in Hawaii, 4.4 in Utah, and 5.8 in Louisiana). Similarly, Baja California Sur and Guanajuato had age-standardized (using 2000 Mexico population aged ≥20 years) diabetes MCD rates of 16.0 and 16.6 per 10,000, but age-standardized diabetes UCD rates varied by a factor of 1.6 (8.0 in Baja California Sur and 13.0 in Guanajuato).

In-hospital deaths, for which more clinical information may be available, had substantially higher probabilities than out-of-hospital deaths of being assigned to cardiovascular diseases (RRR 1.16 in the U.S. and 1.87 in Mexico) (Table 2). In-hospital deaths also had a higher probability of assignment to communicable diseases (RRR 2.89 in the U.S. and 2.84 in Mexico), possibly due to sepsis deaths. Assignment to other noncommunicable diseases was less likely for in-hospital deaths relative to out-of-hospital deaths (RRR 0.91) in the U.S. and more likely (RRR 1.42) in Mexico.

In the U.S., the number of cardiologists was inversely associated with assignment to cardiovascular, other noncommunicable, and communicable diseases, with a 4, 4, and 6% reduction in the probability relative to assignment to diabetes for each additional cardiologist per 100,000 residents, respectively. The probabilities of being assigned to cardiovascular, other noncommunicable, and communicable diseases relative to diabetes increased with additional numbers of endocrinologist and nephrologists (only some of these results were statistically significant). These results may reflect actual or perceived diabetes treatment or additional medical knowledge used in certification, with higher specialized diabetes care.

In the U.S., compared with white men, almost all other race-ethnicity-sex groups had lower probabilities of being assigned to cardiovascular and other noncommunicable causes relative to diabetes. The sex pattern was reversed in Mexico. Decedents with below high school education had a smaller likelihood of being assigned to cardiovascular diseases and other noncommunicable diseases in both the U.S. and Mexico compared with individuals who achieved college or more. There was a higher probability of being assigned to cardiovascular diseases (RRR 1.19) and other noncommunicable diseases (RRR 1.27) and a lower probability of being assigned to communicable diseases (RRR 0.77) if individuals were insured. These findings may all demonstrate differential treatment of diabetes and/or differential availability of information at the time of certification.

The probability of the death of a confirmed diabetic decedent being assigned to cardiovascular disease, relative to diabetes, increased with state BMI and SBP (RRRs 1.03 and 1.01 for each unit change in BMI and SBP, respectively, in the U.S. and 1.20 and 1.01, respectively, in Mexico). This finding probably represents a true epidemiological pattern in which, among those with confirmed diabetes, living in a population with higher BMI and SBP reflects an increased risk of cardiovascular mortality. For deaths assigned to other noncommunicable and communicable diseases, the probability increased with higher BMI in both countries, possibly because of the effects of BMI on noncommunicable diseases as well as general health status. There were no clear effects for SBP for noncardiovascular causes.

After standardizing for the effects of individual- and community-level determinants, the relationship between diabetes UCD and diabetes MCD mortality across states was substantially strengthened in both the U.S. and Mexico (Fig. 2). In the U.S., diabetes UCD rates declined most in Utah, New Mexico, New Jersey, and Louisiana, with the adjusted rates being 32–45% lower than the observed rates. The largest increase in diabetes death rates occurred in Hawaii and California, with 64 and 17% increases in diabetes death rates after adjustment, respectively. In Mexico, adjusted diabetes UCD rates were smaller than those observed in all states, with the largest decreases in Guanajuato, Tabasco, and Puebla (34% for each) and the smallest decreases in Baja California Sur (3%) and Nuevo León (12%).

After standardizing the individual- and community-level determinants to the reference values in Table 1 of the universe of deaths with diabetes as one of the MCD, the number of deaths with diabetes as the UCD declined from 71,276 (33% of diabetes MCD) to 63,874 (30%) in the U.S. (2001) and from 63,607 (67% of diabetes MCD) to 48,318 (51%) in Mexico (2005) (Table 3). Cardiovascular diseases increased from 85,134 (39% of diabetes MCD) to 98,583 (46%) in the U.S. and from 13,070 (14% of diabetes MCD) to 22,539 (24%) in Mexico. There was also an increase for communicable diseases in both countries but the change for other noncommunicable diseases was in the opposite direction for the two countries.

In addition to these direct diabetes deaths, high blood glucose increases the risk of mortality from cardiovascular diseases. By using data on blood glucose from the nationally representative National Health and Nutrition Examination Survey (NHANES) 1999–2002 in the U.S. and the Encuesta Nacional de Salud (National Health Survey [ENSA]) 2000 in Mexico and methods described in detail elsewhere (17), an estimated 106,543 deaths due to IHD (21% of all IHD deaths) and 19,757 deaths due to stroke (12% of all stroke deaths) were attributable to higher-than-optimal blood glucose in the U.S., as were 14,224 deaths due to IHD (26%) and 5,273 deaths due to stroke (19%) in Mexico.

Definitive identification of incomparable coding between diabetes and other diseases among diabetic patients requires detailed prospective studies, with standardized cause-of-death certification. In this analysis, we retrospectively investigated some of the individual and community determinants of cause-of-death assignment with the aim of enhancing the cross-state comparability of diabetes death rates in the U.S. and Mexico.

Standardizing the assignment of diabetes deaths in this analysis to a reference level of individual and community determinants led to significant changes in the diabetes pattern across states and race-sex groups and improved the cross-state comparability of the diabetes cause-of-death statistics. In addition to improved comparability, the reference level of individual and community determinants used in this analysis is one that is expected to provide the most valid cause-of-death statistics; standardizing to this reference level led to a reduction in the number of deaths assigned directly to diabetes in both the U.S. and Mexico.

This study is affected by a number of limitations. The occurrence of a death in the hospital, the number of specialized doctors per capita, and the autopsy variable were valuable proxies for clinical information and diagnostic skills at the community level. Better information on training and diagnostic facilities of the hospitals in which individuals died would allow a more direct assessment of these factors. The death certificate itself may contain errors (e.g., educational attainment and race, which are provided by next of kin). The finding on the role of race, ethnicity, and sex may reflect differences in quality of care experienced by minority patients (18) and differential practices in cause-of-death assignment or actual epidemiological differences in the natural history of diabetes. Finally, our analysis focused on comparability across states in the same country, which limits its application to cross-country comparability.

Above and beyond comparability, improving the overall validity of death certificates is important for public health planning. Improving validity and comparability requires better training in cause-of-death certification, accessible information on medical history (e.g., through linked records), and standardized (and ideally automated) coding. Together with such national efforts, the new revision of the ICD should provide further clarity on the assignment of the UCD as related to diabetes and cardiovascular diseases.