Abstract. Observing stratospheric ozone is essential to assess whether the Montreal Protocol has succeeded in saving the ozone layer by banning ozone depleting substances. Recent studies have reported positive trends, indicating that ozone is recovering in the upper stratosphere at mid-latitudes, but the trend magnitudes differ, and uncertainties are still high. Trends and their uncertainties are influenced by factors such as instrumental drifts, sampling patterns, discontinuities, biases, or short-term anomalies that may all mask a potential ozone recovery. The present study investigates how anomalies, temporal measurement sampling rates, and trend period lengths influence resulting trends. We present an approach for handling suspicious anomalies in trend estimations. For this, we analysed multiple ground-based stratospheric ozone records in central Europe to identify anomalous periods in data from the GROund-based Millimetre-wave Ozone Spectrometer (GROMOS) located in Bern, Switzerland. The detected anomalies were then used to estimate ozone trends from the GROMOS time series by considering the anomalous observations in the regression. We compare our improved GROMOS trend estimate with results derived from the other ground-based ozone records (lidars, ozonesondes, and microwave radiometers), that are all part of the Network for the Detection of Atmospheric Composition Change (NDACC). The data indicate positive trends of 1 % decade−1 to 3 % decade−1 at an altitude of about 39 km (3 hPa), providing a confirmation of ozone recovery in the upper stratosphere in agreement with satellite observations. At lower altitudes, the ground station data show inconsistent trend results, which emphasize the importance of ongoing research on ozone trends in the lower stratosphere. Our presented method of a combined analysis of ground station data provides a useful approach to recognize and to reduce uncertainties in stratospheric ozone trends by considering anomalies in the trend estimation. We conclude that stratospheric trend estimations still need improvement and that our approach provides a tool that can also be useful for other data sets.