[Exploring the applicability of regression kriging in estimating regional food intake across China]

To construct and evaluate the applicability of the regression kriging method for estimating food intake in unsurveyed regions of China, thereby providing method ological support for spatial imputation in national nutrition surveillance.

Using the China Nutrition and Chronic Disease Surveillance(2015) data based on a food frequency questionnaire for adults, a regression kriging was established by combining stepwise regression with ordinary kriging. This method was applied at the county level to spatially impute food intake for different age and sex groups in unsurveyed areas. Leave-one-out cross-validation was used to evaluate model performance using the coefficient of determination(R~2), root mean squared error(RMSE), and mean absolute error(MAE). The differences between estimated and measured food intake values were compared at the national and provincial levels across various population subgroups to assess the method's regional applicability. The result were showed by M(P25, P75).

Covariate analysis showed that gender and average years of education for individuals aged six and above had the greatest influence on adult food intake levels in China, while per capita GDP had the least impact. Among all models, the whole grain model achieved the best fit(R~2=0.479(0.449, 0.553)), and the legume model had the lowest RMSE and MAE with medians of 5.920(5.261, 6.354) and 4.418(3.964, 4.921), respectively. At the national level, the median difference between estimated and observed intake was 2.08%(0.79%, 5.18%) for males and 1.80%(1.46%, 4.76%) for females. Red meat exhibited the smallest estimation error(male: 0.08%, female: 1.42%), followed by whole grains(male: 0.60%, female: 1.42%). At the provincial level, the median difference across all subgroups was 7.21%(2.94%, 14.77%), with the smallest in Tianjin(1.74%(0.75%, 4.17%)) and the largest in Tibet(37.40%(24.23%, 66.64%)).

Regression kriging can be effectively used to impute food intake data in unsurveyed counties during national surveillance. The method performs better for food types with relatively uniform sample distribution and strong spatial autocorrelation of residuals.