In Europe, the Water Framework Directive aims at reaching good ecological status for surface waterbodies by 2015. Consequently European countries are developing methods based on biological (phytoplankton, macroalgae, angiosperms, macrobenthos and fishes), hydromorphological and physico-chemical quality elements for the ecological assessment and monitoring of rivers, lakes and coastal and transitional waters. In these methods, the value of ecological indicators are transformed into ecological status by comparison to the so called “reference conditions”, i.e. the conditions of the indicator in pristine areas (absence of human pressure). Hence setting adequate reference conditions is clearly crucial for the sound assessment of ecological status (Borja et al. 2012). The European Framework 7 project WISER is supporting the implementation of the WFD by testing and complementing existing assessment schemes. The development of suitable methods or the improvement of existing methods for the definition of accurate reference conditions is one aim of WISER. The present work focuses on fish-based quality indicators for estuaries and lagoons (transitional waters under WFD terminology). Fish assemblages highly depend on natural features, both temporal and geographical, at small and large scale. This is especially true in transitional waters where natural abiotic variability is extremely high (Dauvin et Ruellet 2009; Dauvin 2007; Elliott et Quintino 2007). Moreover, the measured indicators (or metrics) characterising fish assemblages highly depend on the sampling method and sampling characteristics. For these reasons, any reference condition for fish in transitional waters must take into account these parameters. There are nearly no transitional waters in Europe that can be considered as being in pristine condition and historical data are not available for all transitional water types. In this context, the aim of the present work is to propose a modelling approach to define type-specific reference conditions for fish assemblages in transitional waters in Europe. The modelling of reference conditions was tested on 13 fish metrics overall, including seven of the most commonly used WFD fish metrics and all the metrics composing the French Estuary and Lagoon Fish Index ELFI. A fish dataset covering 39 estuaries and 14 lagoons distributed across six countries (Bulgaria, Italy, United Kingdom, France, Spain and Portugal) and sampled between 2003 and 2010 was available. For the modelling of reference conditions, two sub-datasets of the best standardized data were selected, one for estuaries and one for lagoons. The dataset for estuaries covers 38 estuaries and contains 1811 trawl hauls. The dataset for lagoons covers 12 lagoons and contains 295 data of fyke net-Cemagref samples collected after a standardised soaking time of about 24 hours. First, selected fish metrics were modelled using Linear Models (LM) and Generalized Linear Models (GLM) taking into account all variables from the sampling protocol, variables from the natural features of estuaries and lagoons and selected pressure indices (Courrat et al. 2009; Delpech et al. 2010; Drouineau et al. 2012), all as fixed effects (Bolker et al. 2009). Pressure indices were calculated from CORINE Land Cover (CLC - Commission of the European Communities, 1994) data 2006 (http://sia.eionet.europa.eu/CLC2006/), except for Stour and Orwell estuary where only CORINE Land Cover 2000 was available. Land cover indices were calculated on three buffers of 1 km, 1.5 km and 2 km around estuaries and lagoons. Second, best LM and GLM were translated in mixed LM (LMM) and mixed GLM (GLMM) with adding a random “estuary” or “lagoon” effect to avoid pseudoreplication in data (Bolker et al. 2009). Last, predictions from the best mixed models were used to define reference conditions. Predictions were made both for a theoretical pristine status (absence of pressure) and for the lowest values of pressure indices observed in the datasets. Preliminary analyses lead to the selection of three pressure indices calculated on a 2 km buffer: the percentage of urban land, the percentage of agricultural land and the percentage of natural land. Only four fish metrics out of the twelve tested for lagoons, and five metrics out of the thirteen tested for estuaries were significantly related to one of the three pressure indices, and effect of pressure indices was always very low (close to zero) in the selected LM and GLM. This may be due to the fact that pressure indices based on land cover and measured on buffers around estuaries and lagoons are not good proxies for human pressure impacting fishes in these transitional waters. When adding a random effect to the LM and GLM, this lead to only three fish metrics responding significantly to some pressure indices. In the end, predictions were computed from the models by linking (respectively) the number of marine migrant species (SR_MM) with the percentage of agricultural land around lagoons, the density of benthic invertebrate feeder fishes (DIB) with the percentage of urban land around lagoons and the percentage of omnivorous individuals (RD_O) with the percentage of natural land around estuaries. All models (including those where fish metrics were not responding significantly to pressure indices) showed a major effect of the sampling method and of natural features of estuaries and lagoons on fish metrics. In particular, the present work argues for the definition of reference conditions specific to sampling gear, sampling season and salinity class. The approach gives interesting results for number of species. For densities and relative densities, results are promising but improvements to the models are required. Results argues for a lack of robustness of the approach for density metrics, especially in the case of fishing events containing very high numbers of fishes, or, in other words, the approach seems very sensitive to outliers. Careful attention should be paid to fish metrics supposed to vary in the same way as pressure, such as RD_O. Indeed, in this particular case, the present approach leads to predicted null relative densities in pristine status.