Urban stormwater runoff is a potential entrance pathway for a wide range of anthropogenic trace pollutants, like biocides, plasticizers, heavy metals or flame retardants, to urban lakes and rivers. However, little is known on dependencies of the occurrence of these trace pollutants on rain event characteristics and climate or seasonal influences. Furthermore, the importance of such dependencies for the calculation of loads and the uncertainties involved are unclear. This thesis evaluates possible correlations between trace pollutant concentrations in urban stormwater runoff and rain event characteristics together with further climate and seasonal influences, based on a large set of measurements from the project “Trace organics in Berlin stormwater runoff (OgRe)”. Here, samples were taken in a one-year monitoring program for five stormwater catchments representing specific urban structure types. Additionally, this thesis investigates whether the consideration of those correlations is necessary for the calculation of loads or whether the use of a mean concentration is sufficient. A method for the correlation analysis is developed and applied to the data, under the requirement to use just one influencing factor (predictor) per correlation and to keep the models simple. Regression models are fitted with regard to normal and log-normal error distributions. The models are evaluated regarding their goodness of fit using the Nash-Sutcliffe efficiency, the log-likelihood ratio, and the prediction coefficient of determination. For 45 out of 48 of the considered substances at least one correlation (i.e. in one of the five catchments) with rain or climate predictors is found. In addition, it is demonstrated that seasonal influences have an effect on substance concentrations for 25 out of 48 substances. Thus, the selected predictor values prove useful to explain the measured concentrations. Only 11 substances show the same correlation with a rain/climate predictor in four catchments and none in all five catchments. So, while concentrations for single events in one catchment can be well explained by the correlations, overall concentration patterns seem to be strongly influenced by the catchment, i.e. its urban structure type. Furthermore, it is shown that the assumption of a normally distributed error does not represent the data adequately in most cases. Consideration of a log-normal error distribution improves most regression models significantly. Regarding single substances, the correlation analysis helps to explain observed patterns. For instance, terbuthylazine, an agricultural pesticide, was only detected during typical application months of May and June, with the same observation in all five catchments. Accordingly, atmospheric deposition from the agricultural surroundings seems a reasonable explanation. In a second example, nicotine was found at very high concentrations in four catchments for low rain event durations, showing a strong decrease with increasing duration. This behavior can be explained by the fast elution of nicotine from cigarette butts within the first minutes of a rain event, followed by dilution during longer rain events. An exemplary load estimation based on a 30-year rain series for Berlin using a Monte Carlo simulation demonstrates that the use of regression models versus mean concentrations can lead to very different results. The reason lies in the selection of sampled rain events which are not distributed according their contribution to the total runoff volume (there should be more small to medium rain events, which contribute more to the total runoff volume). In conclusion, errors in the load estimation can result from i) using a mean concentration instead of a valid correlation, but also from ii) using a non-valid correlation. This underlines the importance of performing a correlation analysis before load calculations, but also the importance of a critical evaluation of the sample data and the correlations. For the latter, a combined evaluation along several goodness-of-fit metrics is suggested, together with plausibility checks of the correlation and of the considered range of values within which the regression model is applied.