20 × 0.20 m frame. Samples were taken and treated following standard guidelines for bottom macrofauna sampling (HELCOM 1988). The occurrence and importance of prey items were inferred from the analysis of fish digestive tracts. The former describes the relative frequency of a particular prey in all digestive tracts, while the latter indicates how much

a particular prey item contributes to the total content in a discrete digestive tract. Both parameters were divided into three categories: high, moderate and low. A ‘high’ occurrence means that a particular benthic animal is found in more than 50% of samples, ‘moderate’ – in 20–50% of samples and ‘low ’ in < 20% of samples. A ‘high’ importance means that most of the KU-60019 in vivo digestive tract can be filled with a particular prey species (more than 50% of tract content), ‘moderate’ – 20–50% of tract content, while‘low ’ means that a particular item is only a small addition to the whole tract content (< 20% of tract content). The occurrence and importance of prey items are shown in Table 1. As the study aimed to evaluate the quality of the seabed for the feeding of fish, the assessment was based only on benthic invertebrates, excluding nectobenthic species and small pelagic fish. To predict the biomass selleck products distribution of prey species the Random forests (RF) regression

model (Breiman 2001) implemented in the ‘randomForest 4.6-2’ package (Liaw & Wiener 2002) within the R environment was chosen. The modelling procedure was as follows. First of all, a correlation matrix was created for all predictors. If a correlation coefficient was > 0.7 or the VIF (variance inflation factors) were > 3, those predictors were not used for constructing

the model. Then the biomass data were split into two sets: train data (70% of all data) for constructing the model and test data (the remaining 30%) for validation. In order to avoid an uneven distribution of zero values the split was made semi-randomly: all sites were chosen randomly but with the proviso that sites with zero values would distribute 70/30 in train/test datasets. Parameters for Org 27569 RF were selected as follows: the number of trees (ntree) was set to 1000, while the number of variables randomly selected at each node (mtry) and minimum node size (ndsize) were set to default values 2.3 and 5 respectively. After running the model the importance of the predictors was assessed. The Mean Decrease Accuracy (%IncMSE) was calculated to assess the importance of every environmental factor for the response variable. During validation, predicted values were compared with observations of external data (test dataset), thereby revealing the model’s true performance. Several estimates were calculated: (1) MAD – mean absolute deviation, (2) CVMAD – coefficient of variation of MAD, rs – Spearman’s correlation between observed (yt  ) and predicted ( y^t) values. equation(1) MAD=n−1∑t=1nyt−y^t, equation(2) CVMAD=ΜADy¯×100.