The surveying and evaluation of the pavement condition provides an objective starting basis to justify decisions in the context of planning maintenance and rehabilitation activities. The survey approaches in Austria, Germany and Switzerland differ i.a. in the type and number of the evaluated distresses and in the length of the measurement sections. A direct comparison between the calculated condition values is feasible only in terms of the rut depth. The evaluation of the distresses follows the same scheme but with different assessments. At first, the evaluated pavement distresses are converted into grades using rating backgrounds and are then combined to sub-indices and to one total condition index providing the basis for the optimization of maintenance treatments. In Switzerland maintenance treatments are assigned manually, directly to the distress types, without the use of an optimization algorithm. The widespread aggregation of different distress types with different causes to one condition value, as well as the combination of different condition values to one total condition erases the underlying failure causes and makes the selection of maintenance treatments considerably more difficult. The common definition of failure criteria (condition thresholds), based on expert opinions or frequency distributions, is not scientifically justified and leads in the case of distress types with no direct relation to road safety to unnecessary loss of service life. It is therefore recommended to limit the use of condition assessments only to visualization of the condition and identification of problem sections. Pavement condition predictions in German speaking countries are almost solely based on empirically derived, deterministic models, whereat the respective performance function (master) is adapted (calibrated) to each road section. The commonly used models however are based on old data, fail to reproduce influencing factors like design parameters, traffic loads and climate conditions in a mathematically complete and correct manner and disregard the presence of censored data. The traditional methods for single-section model calibration are horizontal and vertical shifting, scaling and regression. Predictions with horizontal or vertical shifting of the master performance function through the last measured condition lead however to bias in the initial condition and age distribution. The prediction based on scaling allows, at least theoretically, the reproduction of the performance history. Multiple measurements will not contribute to increasing prediction accuracy, because the shifting resp. scaling of the function is always conducted through the last existing measurement value. If a series of measurements is present, the master function can also be adapted using regression analysis. The derivation of confidence and prediction intervals provides a basis for the transition from deterministic to stochastic prediction at project-level. Ex ante prediction intervals take in consideration the uncertainty in the measurement and prediction of the explanatory variables (e.g. traffic loads). This thesis provides an exemplary calculation of ex ante prediction intervals using bootstrapping und Monte Carlo simulation. The reliability of the prediction increases with the number of the considered measurements and with a decreasing number of free model parameters. The extension of the simple regression model to a multiple regression model through the inclusion of more influencing variables aims to increase the prediction power of the model. Since it is not possible to consider all influencing factors, specification bias is always expected. Furthermore there is a strong correlational relationship (multicollinearity) between cumulative traffic loading und age of the surface layer at project-level. In fact, the existence of multicollinearity does not affect the prediction ability of the model, but it affects the assessment of the contributions of the individual predictors. Homogeneous Markov Chains are a commonly used in engineering literature discrete stochastic prediction model, which requires no knowledge of previous condition development. The performance functions, resulting from the model, exhibit a linear development, irrespective of the particular distress type. The application to real data shows a substantial overestimation of the predicted service life resp. underestimation of the investment needs. Due to the constant transition probabilities, this approach is not applicable to elements with continuous deterioration process. The final comparison of the considered prediction models with a new methodology acc. to HOFFMANN is based on LTPP-data for the distress type transverse cracking and shows a clear picture of the differences in the prediction quality. The methodology acc. to HOFFMANN allows not only the empirical derivation of performance functions and failure distribution, but also accounts for data censoring and provides accurate predictions in comparison.