A hybrid method consisting of data envelopment analysis and machine learning methods

COVID-19, an airborne disease that emerged at the end of 2019 in Wuhan, China, has spread rapidly around the world. Due to the amount and speed of the epidemic, it was declared as a pandemic by the World Health Organization on March 11, 2020. Countries have implemented various strategies in order to both control the contagiousness and increase the effectiveness of the medical treatments applied to the people who have the disease. Evaluation of the results of these strategies has been the main subject of many studies. Data Envelopment Analysis has been widely used in these studies as it is a powerful technique for comparing the performance of decision making units. However, in these studies, scaling the variables according to the population in order to ensure the homogeneity of the decision-making units jeopardizes the consistency of the results. The inconsistency is due to the different population density of each country. The aim of this study is to eliminate the heterogeneity problem in the Data Envelopment Analysis by correctly positioning the effect of population density increase on the transmission and death rate, and to correctly calculate the effectiveness of the medical treatments applied to the people who have the disease in the strategies implemented by the countries to control the contagiousness in the COVID-19 epidemic. In order to ensure the homogeneity of the 85 countries in the study, instead of scaling according to the population, clustering analysis was carried out by using the population densities of the countries. Serial Hierarchical Data Envelopment Analysis, which consists of two different scenarios aiming to measure the infectiousness and medical treatment performance of the countries, was applied to each of the homogeneous clusters obtained as a result of the clustering analysis. It was determined that the efficiency scores obtained as a result of the Serial Hierarchical Data Envelopment Analyzes applied before and after clustering were different from each other.