In semiparametric propensity score analysis, which is a standard in causal inference, we consider incorporating a loss function that is robust to outliers. First, we confirm that the estimation employing covariate balancing is still doubly robust. An over-identified case, in which the number of moment conditions considered in covariate balancing is greater than the number of parameters, is also addressed. Then, we propose a generalized triply robust information criterion, gTRIC, which is valid in this setting. Numerical experiments show that gTRIC performs better than existing information criteria in the presence of outliers.