This paper presents an analysis of red light running (RLR) conducted at saturated intersections in the city of Mumbai, India, where the traffic is highly heterogeneous with respect to vehicle classes and driver behavior. When all vehicles are considered, almost one in 17 drivers is seen to be jumping red signals there. Unlike the RLR behavior that has been previously reported from intersections elsewhere, a peculiarity observed here is that, within a single red phase, two distinguishable segments of RLR behavior exist. The authors classified them into two regimes: Regime 1, just after the onset of red, and Regime 2, just before the onset of the next green. About one-third of RLR events occur in Regime 1 and the rest in Regime 2. The authors fit different distributions on the time distribution of RLR events. The Kolmogorov–Smirnov test suggests that, at all intersections, exponential distribution fits best for RLR behaviors in Regime 1, and extreme value distribution fits for Regime 2. In addition to those two regimes, RLR at a lower rate is observed in the period between those regimes, and normal distribution fits there. To analyze the causal factors of RLR behavior in the two regimes, the authors developed models at a mesoscopic level specific to vehicle class and regime. Although the red-to-green ratio and the presence of policing prove to be relevant factors affecting RLR in both the regimes, the relative time for which the conflict area is free affects RLR in Regime 2 but not in Regime 1.