One of the mechanical learning methods, logistic regression, is applied to the analyses on the rail damage detected on the high rails at the curve sections. The effect of the difference of rail grades on the rail damage severity is estimated to obtain the ideas on the rail grade selection at the curve sections. The logistic regression model with the optimized parameters indicates that an as-rolled rail lowers the rail damage severity compared with a head hardened one on average within the curve radius range from R500 to R800m. From a view point of the damage severity, an as-rolled rail is favorably selected as a high rail in the outer curves with the radius of more than R500 to R600m. Considering the evolution of maximum wear rate of the rail head, a head hardened rail tends to have a little bit smaller damage severity than an as-rolled rail in case the maximum wear rate of an as-rolled rail is quite large. This may attribute to the decrease of the stress loaded by the inhomogeneous worn profile of the rail head.