The performance of bit-interleaved coded modulation in multiple-input multiple-output (BICM-MIMO) systems using an iterative channel estimator is analysed. In a conventional iterative channel estimator, after initialisation with the training phase, the channel estimator switches to the data phase. However, such a conventional iterative channel estimator does not always improve the performance of the receiver. In order to guarantee the performance improvement, a condition on when the output of the decoder should be used by the estimator is determined. Such a condition is related to the reliability of the soft information utilised by the channel estimator. The key in establishing this relationship is to use the mutual information (MI) that the observation vector has about the channel gains given the output of the decoder at each iteration. In this switch-augmented conventional iterative channel estimator, referred to as SAICE, the condition is theoretically found and indicates the needed reliability of the soft information for the channel estimator at the switching time. The switch-augmented scheme guarantees performance improvement of the iterative receiver with each iteration, however, it might need many iterations to converge for moderate to low signal-to-noise ratios (SNRs). A less computationally intensive approach is to use both the training and data segments of the observation. This approach produces a combined iterative channel estimator (CICE) for BICM-MIMO systems. The performance behaviour of the BICM-MIMO system is illustrated through the extrinsic information transfer (EXIT) chart with imperfect channel state information (CSI). Analytical results are verified with computer simulations.