Assessment of Disparities in Diabetes Mortality in Adults in US Rural vs Nonrural Counties, 1999-2018

Key Points Question In the US, do disparities exist in diabetes mortality based on county urbanization? Findings In this cross-sectional study, in 2017-2018 vs 1999-2000, the age-standardized diabetes mortality rates per 100 000 people were unchanged in rural counties (157.2 vs 154.1) but significantly lower in medium-small counties (123.6 vs 133.6) and metro counties (92.9 vs 109.7). In 2017-2018 vs 1999-2000, the mortality rate was significantly higher in rural men (+18.2) but lower in rural women (−14.0). Meaning These findings indicate that overall, US rural counties have persistently high diabetes mortality rates, with additional disparities based on gender.


eMethods 1. Modeling Diabetes Mortality Rates
To model time, we considered the following model fits: 2 time segments (1 change point), 3 time segments (2 change points), and 4 time segments (3 change points). Based on the F-statistic, there was a statistically significant improvement in the model fit in going from '2 time segments' to '3 time segments'. However, there was no statistically significant improvement in going from '3 time segments' to '4 time segments'. Therefore, we used the '3 time segments' fit.
Considering 720 cells (3 urbanization categories × 3 age-groups × 2 sexes × 4 regions × 10 time points) as independent observations, weighted multiple linear regression was used to model the outcome [log (mortality rate)], where mortality was calculated as deaths per population in each cell, and reported as annual diabetes mortality rate per 100,000 people (ADMR). Under the assumption that log (mortality rate) for a given cell was based on the underlying Poisson count variable, the optimal weight for a given data cell would be the expected mortality count for that cell. We used a main effects model (region, county urbanization, age-group, and sex, used as nominal variables, and time as a 3-segment linear function), to estimate optimal weights as predicted mortality counts. The procedure was iterated until it converged, and the resulting weights were used for all subsequent analyses. Although the Poisson assumption was used to provide weights, standard errors of parameter estimates were based on the model, i.e., they used the residual mean squared error (MSE).
Weighted regression was used on the assumption that the 720 data points (counts of diabetes related deaths and population denominators) have a Poisson variation. However, the Poisson assumption may overstate the precision. Therefore, we modeled the log of the rate (counts of diabetes related deaths/population) in a general linear model, maintaining the consequence of the Poisson assumption that the cells have different precision. Therefore, we used weights proportional to the expected counts. This resulted in substantially smaller standard errors than those for unweighted regression.
The model for log (mortality rate) was based on consideration of all main effects and up to 3-way interactions. Interaction terms were pruned based on the statistical significance. The hierarchical structure was maintained such that inclusion of a high-order interaction required inclusion of corresponding lower-order interaction terms. eMethods 2. Unadjusted and Adjusted Estimates for Diabetes Mortality To determine unadjusted estimates, we added predicted death counts and population denominators for each combination of variables, along with age-group, to obtain mortality rates for table cell (for each age-group). These rates were age-standardized to the 2009-2010 population by taking weighted means across age-groups, using weights determined by the 2009-2010 age distribution within our dataset.
To determine adjusted estimates, we created and appended copies of the original dataset without the Y-variable [log (mortality rate)], one for each cell, such that the non-tabled variables had the distribution of the original dataset with the variables fixed to a particular set of values. These counts were added over the non-tabled variable for each combination of table value and agegroup, with the same age adjustment as for the unadjusted tables. The predicted log (mortality rate) was calculated for each combination.

eMethods 3. Comparison of Simple and Full Models
To examine if variation in mortality time trends with respect to other exposures (i.e., sex, age-group, and region) accounted for the variation in mortality time trends across urbanization categories, we compared F-statistics from two sets of models: simple and full. The 'simple' set compared a model with no time*exposure interactions with one containing only the time*urbanization interactions (6 numerator d.f.), which had an F-statistic of 34.0. The 'full' set was based on adding time*urbanization interactions to a model with all other time*X interactions (6 numerator d.f.), which had an F-statistic of 31.1. Both models contained all main effects and all 2-way interactions other than those involving time. The same denominator mean square based on 661 d.f. was used in both cases. The similar F-statistic values suggested that almost all variation in time trends due to urbanization was independent of variation in time trends with respect to other exposures.  -group, years 1999-2000 2003-2004 2009-2010 2017-2018 1999-2000 2003-2004 2009-2010 2017-2018   Estimated annual diabetes mortality rates (ADMR [95% confidence interval]; per 100,000 people) from unadjusted and adjusted estimates that included sex, region, and up to 3-way interaction terms (age-group, county urbanization, year). Statistical significance at *P<.01; **P<.001; ***P<.0001 compared to 2017-2018.