3. Results
Figure 1 and Figure 2 show that individuals living in southern and eastern coastal regions in China have greater longevity. Longevity is lower in northern regions, which feature lower humidity, higher monthly mean temperature standard deviation, and lower Se. Spearman’s rank correlation was used to evaluate the relationship between age and element concentration (Table 2). Positive correlations were seen between longevity ratio and sea fish consumption, humidity degree, T-Se content in soil, WS-Se content in soil, and freshwater food consumption, while negative correlations were seen for longevity ratio and altitude, monthly mean temperature standard deviation and meat consumption. However, meat consumption was not statistically as significant as the other seven factors. A regression analysis of these factors and longevity calculated associated correlations, as illustrated in Table 3.
Table 3 shows correlation coefficients for the nine factors in the regression model. Sea fish consumption, humidity, T-Se, altitude, and standard deviation of monthly mean temperature were more closely related to longevity than WS-Se, fresh water food consumption, or meat consumption. p values associated with meat consumption were higher than 0.05 in 2000 and 0.01 in 2010, indicating that neither factor was significant. No R2 was higher than 0.4, indicating that longevity is not determined by one single factor, and that a multiple factor analysis was necessary. Stepwise MLR was used to ascertain the association between factors and longevity ratio, Table 4 and Table 5 represent the stepwise MLR developed using longevity ratio as the dependent variable and environment and nutrition factors as the independent variables in 2000 and 2010, the R2 is 0.535 and 0.543, respectively. The final three independent variables of the model were standard deviation of monthly mean temperature, altitude and per capita sea fish consumption, both in 2000 and 2010.
After modeling, the classical multiple regression model fit was always lower than 0.6, and the Durbin-Watson value was 1.268. This showed that the residual of ordinary least square (OLS) has a positive autocorrelation structure and was not suitable for establishing models with OLS. The regression parameters in this study varied relative to geographic location. Regression parameter estimation of a global regression model is the average value of the regression parameters in a study area. This fails to reflect the spatial character of all the regression parameters. Factors in this study such as Se, temperature, humidity, altitude and fish consumption vary according to regional characteristics and spatial effects. In this case, we adopted a geographically weighted regression (GWR) model. In GWR, geographic location was embedded in the regression parameters. Consequently, observational data close to the position has more influence than data far from the position, resulting in adjacent locations having similar regression parameters. We tested all the factor groups and found that the correlation coefficient of the group consisting of humidity, altitude, and per capita sea fish consumption was highest, both in 2000 and 2010. In addition, the top four groups with highest R2 are illustrated in Table 6 and Table 7.
In 2000 and 2010, regression coefficient of humidity, altitude and per capita sea fish consumption were always higher than groups made up of other factors, which indicates that the three factors are most related to longevity at the city scale. Regression coefficient of WS-Se, altitude and per capita sea fish consumption were always ranking the second place.
By comparing Table 4 and Table 6, and by comparing Table 5 and Table 7, the correlation coefficient of each group significantly improved (higher than the multiple liner regression model) in the GWR model. The predicted regression value of longevity by GWR with humidity, altitude, and per capita sea fish consumption, for each city, are illustrated in Figure 2.
By comparing Figure 13 and Figure 1, and by comparing Figure 14 and Figure 2, we observe a good fit with most places in China, especially in the middle and eastern regions. The relation between the real longevity index and the simulated longevity index is illustrated in Figure 15 and Figure 16.