Background and purpose: The purpose of this large multicentre prospective cohort study was to identify which dose volume histogram parameters and pre-treatment factors are most important to predict physician-rated and patient-rated radiation-induced swallowing dysfunction (RISD) in order to develop predictive models for RISD after curative (chemo) radiotherapy ((CH) RT). Material and methods: The study population consisted of 354 consecutive head and neck cancer patients treated with (CH) RT. The primary endpoint was grade 2 or more swallowing dysfunction according to the RTOG/EORTC late radiation morbidity scoring criteria at 6 months after (CH) RT. The secondary endpoints were patient-rated swallowing complaints as assessed with the EORTC QLQ-H&N35 questionnaire. To select the most predictive variables a multivariate logistic regression analysis with bootstrapping was used. Results: At 6 months after (CH) RT the bootstrapping procedure revealed that a model based on the mean dose to the superior pharyngeal constrictor muscle (PCM) and mean dose to the supraglottic larynx was most predictive. For the secondary endpoints different predictive models were found: for problems with swallowing liquids the most predictive factors were the mean dose to the supraglottic larynx and radiation technique (3D-CRT versus IMRT). For problems with swallowing soft food the mean dose to the middle PCM, age (18-65 versus >65 years), tumour site (naso/oropharynx versus other sites) and radiation technique (3D-CRT versus IMRT) were the most predictive factors. For problems with swallowing solid food the most predictive factors were the mean dose to the superior PCM, the mean dose to the supraglottic larynx and age (18-65 versus >65 years). And for choking when swallowing the V60 of the oesophageal inlet muscle and the mean dose to the supraglottic larynx were the most predictive factors. Conclusions: Physician-rated and patient-rated RISD in head and neck cancer patients treated with (CH) RT cannot be predicted with univariate relationships between the dose distribution in a single organ at risk and an endpoint. Separate predictive models are needed for different endpoints and factors other than dose volume histogram parameters are important as well. © 2011 Elsevier Ireland Ltd. All rights reserved.