The Long Term Evolution (LTE) link adaptation is based on measured instantaneous Signal to Interference and Noise Ratio (SINR) which is used for selecting Modulation and Coding Scheme (MCS) for transmissions. In addition, depending on the scheduler, SINR may be used to determine which users are scheduled for a certain transmission time interval and on which frequency resources. The measured SINR can be inaccurate due to measurement errors, rounding errors due to quantization of the SINR values, and delay from time of measurement until the actual data transmissions. To compensate for SINR inaccuracies, the SINR can be adjusted by a certain offset before being used for link adaptation and scheduling. This offset value, referred to as the link adaptation margin in this thesis, can be a fixed value common to all the users in the system at all times or adaptively adjusted for each user based on some algorithm via a feedback loop, referred to as differentiated link adaptation. This thesis tries to improve the system performance for LTE downlink and uplink by using differentiated link adaptation based on packet error occurrences of each user as feedback. The performance of the differentiated link adaptation was compared to the best performance that is achievable using a fixed link adaptation margin. We investigated the influence of several parameters on the link adaptation error characteristics, such as settings for SINR estimation, scheduling algorithms, traffic patterns. It was shown that there are error clusters, but that these are short and difficult to react to on time. A performance gain was only possible in the downlink for FTP traffic with a proportional fair in time and frequency (PFTF) scheduler which was the scenario with the largest variations with regards to both scheduling and traffic model. It was seen that the gains of using differentiated link adaptation increased in the downlink as the transmissions got more random. For more stable situations, a fixed link adaptation margin performed better. The uplink performance was worse with differentiated link adaptation than with a fixed optimized link adaptation margin. This could be because the uplink SINR estimation was much better than in the downlink, with low estimation error variance, in which case frequent SINR adjustments could make the situation worse off.