Analysis of atmospheric occultation data from the GPS Meteorology experiment has revealed that the refractivity retrievals in the lower troposphere were systematically smaller than those obtained with numerical weather prediction models. It has been suggested that the bias was due to a combination of atmospheric multipath, critical refraction, and receiver tracking errors. In this paper, we show that a similar bias exists in the CHAMP and SAC-C data and describe the characteristics of the bias based on over 6700 soundings from October 2001. Retrievals obtained using the recently introduced canonical transform method are shown to markedly reduce the refractivity bias; however, a significant bias still remains below 2 km altitude. To better understand the underlying causes of the bias, we perform an end-to-end simulation study that incorporates full-wave signal propagation and realistic receiver tracking effects using an ensemble of atmospheric profiles. We find that atmospheric ducting effects associated with the top of the planetary boundary layer (PBL) at 1--2 km altitude would cause retrieval errors at and below the PBL even in the absence of the receiver errors. Furthermore, current implementation of the receiver tracking algorithm based on an enhanced version of the phase-locked loop could introduce additional errors under the low signal-to-noise ratio conditions that are often encountered in the lower troposphere. The latter problem is expected to be resolved in the near future through the adoption of open-loop tracking and the removal of the navigation modulation from the GPS signal.