This thesis focuses on the development and evaluation of prediction models for knee and low back pain in primary care. Knee complaints as well as low back pain are common musculoskeletal disorders in the western countries and patients with these complaints are frequently treated by physical therapists in primary care. It is known that persistent complaints, in particular, put considerable pressure on society, as well in terms of costs for the employer and employee as in terms of quality of life for patients and their families. Conservative treatment by physical therapists can possibly prevent or delay surgery or other interventions, which may reduce health care costs. Neither physical therapy nor other treatment strategies are standardized, and at the same time, the effects of various treatment strategies are small. Several researchers examined the effects of physical therapy treatment and concluded that the evidence for effects in patients with knee or back complaints is small. This is probably due to the large heterogeneity in patients, type of interventions and different outcome measures used. Prediction models are needed to distinguish between high and low risk patients and can be helpful for physical therapists in providing a correct prognosis and an individual treatment plan. Knowledge about the clinical course and possible prognostic factors is essential in order to make individualized treatment decisions. Therefore, more insight is needed in which prediction models for musculoskeletal complaints (and especially for knee and back complaints) are usable yet in primary care, which models can be useful after validation and what methods can be used to make developing high quality prediction models easier in clinical research. Therefore, in chapter two a systematic review has been carried out according to existing prediction models for patients with Anterior Knee Pain in primary care to examine if there are models that are directly usable in daily practice or that some existing models can be updated. In chapter three a systematic review has been carried out according to existing prediction models for patients with nontraumatic knee pain in primary care with the same objective and to examen if new summary prediction models can be developed based on the existing models and internally and externally validated. To gain a well-defined outcome measure, in chapter four, LCGA has been used to define a dichotomous outcome measure that leads to a stable and validated prediction model for LBP. In chapter five of this thesis, we evaluated if a former predictor variable for chronic LBP is still clinically relevant in the light of the novel method for evaluating the discriminative performance of a model, the Decision Curve Analysis (DCA) using the Net Benefit (NB). This method is able to identify the number of patients that are better classified and permits an understanding of the clinical value of methods for patient selection by prediction models. Last but not least, a common problem in developing prediction models is the amount of missing data in the available data sets. Multiple Imputation (MI) is the recommended method for processing these incomplete data. If a dataset contains missing values, it is recommended to apply MI before excluding variables in logistic regression with backward selection (BWS) from the pooled model. In chapter six, the objective was to evaluate the results of four different pooling methods (D1, D2, D3 and the Median-P-Rule) in Multiply Imputed datasets after a BWS-procedure in simulated datasets and compare these with the results in the complete dataset (without missing data). All analyzes were repeated in real-world datasets to be sure of the usability in daily practice. Especially, the stability of the developed prediction models was evaluated.