Performance-based earthquake engineering (PBEE) has in many ways revolutionized the thinking about seismic engineering design and acceptable performance of buildings in earthquakes. It is now making its way into commercial engineering design and risk analysis practice, as engineers aim to design better-performing buildings, and holders of mortgage or insurance instruments try to better understand the risk they face from damage to associated buildings. Some parts of the calculations (e.g. structural response) have been extensively assessed and validated. There are few similar studies, however, that focus on the damage and loss predictions. The purpose of this dissertation is to address this, by analyzing, evaluating, and improving the damage and loss predictions. The specific PBEE methodology examined in this dissertation is the FEMA P-58 Seismic Performance Assessment Procedure. FEMA P-58 damage and loss predictions are analyzed, to determine how they are impacted by other parts of the calculations. Firstly, variance-based sensitivity analyses are conducted to investigate the interaction of loss predictions with different inputs to the calculations. Of the six inputs considered in the analyses, it is found that predictions of building repair cost (as a fraction of replacement value) are most sensitive to shaking intensity and building age, while building re-occupancy time predictions are most sensitive to shaking intensity and building lateral system. Secondly, a methodology is developed to quantify the impact of available structural response data from seismic instrumentation on the quality of the damage and loss predictions. The density of instrumentation examined using the methodology ranges from the case in which all floors are instrumented to that in which no floors are instrumented and simplified procedures are used to produce structural response predictions. It is found that the quality of the predictions generally improves as the density of seismic instrumentation increases, but it is not crucial for the density to be very high to achieve reasonable accuracy in both damage and loss predictions (although this may depend on the arrangement of instrumentation within a building). Loss predictions are evaluated using data observed in previous seismic events, to understand the degree to which they reflect real-life consequences of earthquakes. A methodology is developed for evaluating the ability of FEMA P-58 component-level losses to predict damage observed for groups of buildings. It is found in applications of the methodology that FEMA P-58 non-structural component-level loss predictions provide more insight into damage than variations in ground shaking between buildings. Finally, this dissertation includes a number of recommendations for improving non-structural mechanical component fragility functions and associated loss predictions used in FEMA-58 calculations. The fitting technique currently used for the functions does not converge in some cases, and the methodology used to predict anchored mechanical component losses can lead to some unexpected results, such as non-smooth variation of repair costs with anchorage capacity. An alternative statistical technique is proposed for fitting the fragility functions that mitigates the non-convergence problems when fitting and makes predictions that better align with damage observed in past events. A more intuitive methodology for predicting anchored mechanical component losses is also suggested. The findings of this dissertation help to enhance understanding of, and improve, the damage and loss predictions used in the FEMA P-58 seismic performance assessment procedure. They ultimately enable various stakeholders, such as building owners, design professionals, lenders, and insurers, to make more informed decisions about seismic risk.