Demand Response (DR) strategies represent an innovative option to optimise energy management. In particular, smart thermostats have captured the attention of the scientific community for their effectiveness in achieving energy-saving and peak-shaving by lowering HVAC consumption during critical hours of the year. One way of achieving this aim, is to leave the control of the smart thermostat to a third party for the duration of the DR event in the so-called Direct Load Control (DLC) configuration. Most research focuses on thermostat overrides during DR events; in this work, we use real world data from the Donate Your Data dataset to analyse the interaction of users with the thermostat around the DR event. In particular, this work focuses on users that interact with the thermostat before (anticipative behaviour) or during the DR event (reactive behaviour), leading to a lower efficiency of the load control. Through clustering techniques, different categories of users are identified, and some significant cases are simulated on a building energy simulation tool to quantify the missed power reduction and the impact on energy. The study highlights that the behaviour of some users can reduce or even nullify the efficacy of the DLC strategy. In light of the findings and to prevent this issue, we suggest the need for tailored DR events for different archetypes of users as identified in this work through clustering.