New algorithm for constructing area-based index with geographical heterogeneities and variable selection: An application to gastric cancer screening

To optimally allocate health resources, policy planners require an indicator reflecting the inequality. Currently, health inequalities are frequently measured by area-based indices. However, methodologies for constructing the indices have been hampered by two difficulties: 1) incorporating the geographical relationship into the model and 2) selecting appropriate variables from the high-dimensional census data. Here, we constructed a new area-based health coverage index using the geographical information and a variable selection procedure with the example of gastric cancer. We also characterized the geographical distribution of health inequality in Japan. To construct the index, we proposed a methodology of a geographically weighted logistic lasso model. We adopted a geographical kernel and selected the optimal bandwidth and the regularization parameters by a two-stage algorithm. Sensitivity was checked by correlation to several cancer mortalities/screening rates. Lastly, we mapped the current distribution of health inequality in Japan and detected unique predictors at sampled locations. The interquartile range of the index was 0.0001 to 0.354 (mean: 0.178, SD: 0.109). The selections from 91 candidate variables in Japanese census data showed regional heterogeneities (median number of selected variables: 29). Our index was more correlated to cancer mortalities/screening rates than previous index and revealed several geographical clusters with unique predictors.