Comparing methods of measuring geographic patterns in temporal trends: an application to county-level heart disease mortality in the United States, 1973 to 2010

Abstract

Purpose: To demonstrate the implications of choosing analytical methods for quantifying spatiotemporal trends, we compare the assumptions, implementation, and outcomes of popular methods using county-level heart disease mortality in the United States between 1973 and 2010.

Methods: We applied four regression-based approaches (joinpoint regression, both aspatial and spatial generalized linear mixed models, and Bayesian space-time model) and compared resulting inferences for geographic patterns of local estimates of annual percent change and associated uncertainty.

Results: The average local percent change in heart disease mortality from each method was -4.5%, with the Bayesian model having the smallest range of values. The associated uncertainty in percent change differed markedly across the methods, with the Bayesian space-time model producing the narrowest range of variance (0.0-0.8). The geographic pattern of percent change was consistent across methods with smaller declines in the South Central United States and larger declines in the Northeast and Midwest. However, the geographic patterns of uncertainty differed markedly between methods.

Conclusions: The similarity of results, including geographic patterns, for magnitude of percent change across these methods validates the underlying spatial pattern of declines in heart disease mortality. However, marked differences in degree of uncertainty indicate that Bayesian modeling offers substantially more precise estimates.

Keywords: Bayesian methods; Generalized linear mixed models; Geographic patterns; Heart disease mortality; Joinpoint regression; Spatial analysis; Spatiotemporal trends.

Figure 1. Distributions of county-level age-adjusted heart disease death rates, per 100,000, ages 35 and older, 1973–2010

Plot shows median, 25^th and 75^th percentile rates as a box. Whiskers represent 1.5 IQRs above and below the 25^th and 75^th percentile, respectively. The y-axis has been truncated to remove extreme outliers. (n=3,099). Rates are age adjusted to the US standard 2000 population.

Figure 2. Estimated percent change for heart disease death rates by method, New Jersey and Montana, 1973–2010

Graphs compare estimates of biennial percent change in heart disease death rates for a single county across selected pairs of analytic methods. Horizontal lines represent the 95% confidence interval for the method listed on the x-axis. The black line represents the identity line (i.e. where the two methods estimate the same value). This figure is provided for these two states as an illustration of the results of these methods using states that generally contain counties with large and small populations (New Jersey and Montana, respectively). Graphs of comparisons between joinpoint regression and each method for these two states are provided in Supplemental Figure B. Note that the range of the x-axis is consistent for all graphs within a state, but not across states, and the y-axis is consistent across all graphs.

Figure 3. County-level estimated biennial percent change in heart disease death rates, United States, 1973–2010, by method

Estimates calculated using (A) joinpoint regression, (B) aspatial GLMM, (C) spatial GLMM, and (D) Bayesian space-time model. (Note that this figure is for color reproduction.)

Figure 4. Variance of county-level estimated biennial percent change in heart disease death rates, United States, 1973–2010, by method

Variance estimated by (A) joinpoint regression, (B) aspatial GLMM, (C) spatial GLMM, (D) Bayesian space-time model. (Note that this figure is for color reproduction.)