In this work we explore the performance of a selective decision directed (DD) channel estimation (CE) method for orthogonal frequency division multiplexing (OFDM) based communications over multipath Rician fading channels, where fading is modeled as an auto-regressive (AR) random process. The method is based on soft-selection of data sub-carriers that are used in the DD mode. Particularly, in the process of soft-selection we weigh the data sub-carriers for reliable CE, and couple them with DD CE based on least-squares (LS) with thresholding. Our numerical results indicate performance improvements as compared to block-by-block pilot-assisted CE with a uniform pilot grid, and non-weighted DD CE. For uncoded and low density parity check (LDPC)-coded OFDM systems, we explore the trade-offs resulting from using a varying number of pilots and/or DD data sub-carriers, the effect of the receiver-end diversity, and the impact of channel properties (such as the Rician k-factor and the Doppler spread) in terms of the overall system throughput and bit error rate (BER).