As the average life expectancy increases, particularly in developing countries, prevalence of neurodegenerative diseases has also increased. This trend is especially alarming for Alzheimer's disease (AD); as there is no cure to stop or reverse the effects of AD. However, recent pharmacological advances can slow the progression of AD, but only if AD is diagnosed at early stages. We have previously introduced an ensemble of classifiers based approach for combining event related potentials obtained from different electrode locations as an effective approach for early diagnosis of AD. We further expand this approach and analyze its robustness and stability in two ways: comparing the diagnostic accuracy on hand selected and cleaned data vs. standard automated preprocessing, but more importantly, comparing the diagnostic accuracy on two different cohorts, whose data are collected under different settings: a research university lab and a community clinic.