Previous studies have identified several important factors r
Previous studies have identified several important factors related to CRC mortality. Socioeconomic disparities and area-level socioeconomic disadvantage have been found to associate with CRC mortality [, , , ]. Better access to primary care and living in urban areas have also been shown to be associated with lower CRC mortality [, , , ]. Some behavioral risk factors such as CRC screening, cigarette smoking, and overweight have also been demonstrated to be important predictors of CRC mortality [, , , , , , , ].
The findings from these existing studies shed some light on determinants related to CRC mortality, but they may not appropriately explain the geographic variations because the studies did not take into consideration the underlying spatial structure. Examining determinants of geographic variation for CRC mortality is complicated by the existence of hotspots of CRC mortality, as it AngiotensinI indicates the rates do not distribute randomly across space. Analysis ignoring such spatial dependence often produce inaccurate results and invalid inference [8,, , ]. In this study, we aimed to address spatial autocorrelation using spatial econometric models to obtain unbiased estimates for the association between CRC mortality and likely county-level determinants within NC.
Results The mean county-level age-adjusted CRC mortality rate was 45.1 per 100,000 population, ranging from 29.4 to 75.8 (Table 2). Spatial distribution of CRC mortality rates showed higher rates in the northeastern areas of the state compared to the other counties (Fig. 1). Global Moran’s I was highly significant for spatial autocorrelation (Moran’s I = 0.2897, p < 0.001), indicating that neighboring counties have similar values. In addition, the LISA map identified the location of high-high clusters in the northeastern counties (Fig. 2). Table 2 presents summary statistics for all variables included in the analysis. The proportion of people within a county without CRC screening ranged from 18.1% to 55.7%. The Z scores for ACSC admissions, obesity, and cigarette smoking ranged from -2 to +3. The Z scores for SES deprivation were between -0.7 and + 0.7 across counties. The median proportion for race and ethnicity was highest for non-Hispanic black population (18.6%) and lowest for Native Americans (0.4%). Overall, 19% of the population under 65 did not have health insurance. Seventy counties were in an HPSA area for primary care while 40 counties were urban and 41 were small town/rural areas. There were 35 counties in the low CRC risk group and 28 counties in the high risk group. Table 3 presents the results from both the OLS model (non-spatial estimates) and spatial lag model (spatial estimates). Overall, in general, the estimates in the spatial lag model had smaller effects than those in the OLS models. Because the spatial estimates are more accurate than the non-spatial estimates due to the existence of spatial dependence, we focused on the interpretation of the spatial estimates. Significant (p < 0.05) results were found for SES deprivation, interaction terms between SES deprivation and risk groups, and large town counties in the spatial lag model. To interpret the effects of spatial lag estimates, one needs to keep in mind that the CRC mortality rate for a specific county is affected by the independent variables in the same county (direct effect, estimated by β), and by the independent variables in the neighboring counties and their neighbors (indirect effects). The average total effect can be estimated using formula β/(1-ρ) [34,35]. For example, on average, the direct effect of large town counties was 4.116 CRC deaths more per 100,000 population compared to urban counties (Table 3). Given the spatial dependence, the average total effect for a county being a large town was associated with 11.79 more CRC deaths per 100,000 population than that of urban counties.