The significant interaction effects between SES deprivation
The significant interaction effects between SES deprivation and risk group make the interpretations for the main effects less meaningful, as the effect of SES deprivation differed by risk group. To directly assess and compare the SES deprivation effect by risk group, we computed the slopes of the SES deprivation effect to estimate the average total effect of SES deprivation for each risk group: 60.78, 0.76 and 1.32, for low, moderate and high risk groups. This indicates that a one standard deviation increase in SES deprivation was associated with, on average, an increase of 61 CRC deaths per 100,000 Dalbavancin for the low risk group, and with a negligible effect in areas where behavioral risk was moderate or high. To better demonstrate these differential effects, we computed the model predicted CRC mortality rates and plotted the prediction by risk group for selected SES deprivation Z scores. Fig. 3 shows that as SES deprivation increased the predicted CRC mortality rates increased, with larger increases in low risk group than moderate and high risk groups (with an increase of 11 people per 100,000 versus 3 and 6, respectively, from SES deprivation Z score -0.25 to 0.25). In addition, the CRC mortality rates were smaller for low risk group as compared to the other risk groups in less SES deprived areas (39.4 for low risk group versus 45.5 and 42.7 for moderate and high risk groups where SES is -0.25). Furthermore, the difference in rates between the risk groups decreased as SES deprivation increased (compared to the data where SES is -0.25, the rate is 50.5 for low risk group versus 48.7 and 49.0 for moderate and high risk groups where SES is 0.25).
Discussion Using NC county-level mortality data from 2003 to 2013, we identified clusters of counties with high CRC mortality in the northeastern area of the state. This finding is consistent with a recent analysis using national data from 2000 to 2009 . We also demonstrated that estimates differed between an OLS model and a spatial lag model, which underscores the importance of accounting for spatial process when the data reveal spatial autocorrelation. This finding confirmed the evidence found in previous studies [8,, , ]. Furthermore, our finding suggests a spatial diffusion process for the CRC mortality rates across NC counties; CRC mortality in one county influenced CRC mortality in the neighboring counties. This process likely worked through shared healthcare resources and informal interactions among individuals in the neighboring areas. In this study we found that the CRC mortality rates were lower in urban counties than that in counties with large non-metropolitan towns, but the rates did not differ between urban and small town/rural counties. This finding is partially consistent with prior studies, in which lower CRC mortality rates were seen in urban counties than both less urban and rural counties [7,10]. It is not clear why we did not find a difference in CRC mortality between urban and small town/rural counties in the regression models. Additional model (data not shown) grouping the non-metro counties into non-metro urban and rural counties and still found no difference between urban and rural counties. Future studies will need to further examine this issue. Because area-level factors are geographically/regionally specific, our findings may not generalize to other geographic areas. Another limitation of our study is the lack of person-level and tumor level data. This precludes comprehensive analyses to assess the effects from an individual level and the dynamic processes from interactions between individuals and the environment. Unfortunately, there is no state or national data source with complete data on key determinants at both an individual and area-level to allow us to assess geographic variation of CRC mortality. We recognize that without complete data at tumor, person, and area levels, our study is unable to fully distinguish whether the geographic variation is from the characteristics of areas (i.e., contextual) or from the types of individuals living in these areas (i.e., compositional) [36,37]. Until more data become available, a spatial approach that correctly accounts for spatial dependence remains appropriate and valuable to identify determinants at aggregate level to inform population level interventions.