A large systematic di erence (ranging from − cm to + cm) was found between NAVD 88 (North American Vertical Datum of 1988) and the pure gravimetric geoid models. This di erence not only makes it very dicult to augment the local geoid model by directly using the vast NAVD 88 network with state-of-the-art technologies recently developed in geodesy, but also limits the ability of researchers to e ectively demonstrate the geoid model improvements on the NAVD 88 network. Here, both conventional regression analyses based on various prede ned basis functions such as polynomials, B-splines, and Legendre functions and the Latent Variable Analysis (LVA) such as the Factor Analysis (FA) are used to analyze the systematic di erence. Besides giving a mathematical model, the regression results do not reveal a great deal about the physical reasons that caused the large di erences in NAVD 88, which may be of interest to various researchers. Furthermore, there is still a signi cant amount of no-Gaussian signals left in the residuals of the conventional regression models. On the other side, the FA method not only provides a better t of the data, but also o ers possible explanations of the error sources. Without requiring extra hypothesis tests on the model coe cients, the results from FA are more e cient in terms of capturing the systematic di erence. Furthermore, without using a covariance model, a novel interpolating method based on the relationship between the loading matrix and the factor scores is developed for predictive purposes. The prediction error analysis shows that about 3-7 cm precision is expected in NAVD 88 after removing the systematic di erence.