Ribosome profiling provides a detailed view into the complex dynamics of translation. Although the precise relation between the observed ribosome footprint densities and the actual translation elongation rates remains elusive, the data clearly suggest that elongation speed is quite heterogeneous along the transcript. Previous studies have shown that elongation is locally regulated by multiple factors, but the observed heterogeneity remains only partially explained. To dissect quantitatively the different determinants of translation speed, we here use probabilistic modeling of the translation dynamics to estimate transcript-specific initiation and local elongation rates from ribosome profiling data. Using this model-based approach, we estimate the fraction of ribosomes (~9%) undetected by the current ribosome profiling protocol. These missing ribosomes come from regions harboring two or more closely-stacked ribosomes, and not accounting for them leads to a substantial underestimation of translation efficiency for highly occupied transcripts. We further quantify the extent of transcript- and position-specific interference between ribosomes on the same transcript, and infer that the movement of ~2.5% of ribosomes is obstructed on average, with substantial variation across different genes. The extent of interference also varies noticeably along the transcript sequence, with a moderately elevated level near the start site and a significantly pronounced amount near the termination site. However, we show that neither ribosomal interference nor the distribution of slow codons is sufficient to explain the observed level of variation in the mean elongation rate across the transcript sequence. Surprisingly, by optimizing the fit of statistical linear models, we find that the hydropathy of the nascent polypeptide segment within the ribosome plays a major role in governing the variation of the mean elongation rate along the transcript. In addition, we find that positively and negatively charged amino acid residues near the beginning and end of the ribosomal exit tunnel, respectively, are important determinants of translation speed, and we argue that this result is consistent with the known biophysical properties of the exit tunnel.