We propose a novel approach for estimating a logistic regression model with error-prone predictors. Particularly, we consider the case where the predictors are subject-specific intercepts/slopes from a longitudinal random effects model, in a high dimensional setting. The model is optimized under conditioning on sufficient statistics for the latent subject specific predictors, hence the estimation is free of error-prone predictors. To handle high dimensional longitudinal predictors, we employ longitudinal principal component analysis. To achieve stability in estimation, a robust majorization on the negative conditional log-likelihood is applied. The method outperformed naive estimates in our numerical experiments. The method is applied to cortical thickness data to predict dementia transition from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) data. This is a joint work with Hyung Park.