This study examines the attentional blink (AB) under conditions where attention was directed either to the overall (global) or to the component (local) form of compound Navon letters. Forty-five adult participants were presented RSVP sequences of compound letter stimuli and were asked to identify the prescribed attentional aspect of a red target letter (T1) and to detect the presence of a second item (T2 = letter X), again of specified global or local aspect. Trials were presented at either four or six letter forms per second (f/s). We report that the AB estimated from the mean performance across participants is much longer (1.5-2.5 s) than that reported previously with single letter RSVP (200-500 ms). A lag-1 sparing effect was not observed when data were filtered for correct identification of T1. Characteristic rapid (initial) and slower (later) recovery phases suggested the use of a bilinear ("knee") recovery model, possibly identifying a novel slow recovery mechanism for the AB when its duration is sufficiently long. However, it is proposed that the duration of the AB should be analyzed by estimation of each individual's AB prior to group analysis. Analysis of individual bilinear fits to the data shows a significant effect of attentional condition for the 4 f/s data, with more rapid recovery for Global-Global (GG) than Local-Local (LL) target conditions, consistent with the global precedence hypothesis. While the 6 f/s data did not demonstrate such differences between attentional conditions, it was at the expense of T1 identification performance with significantly poorer T1 identification for the LL cf GG conditions. Individual fits also demonstrate that the later slow recovery phase seen in the population mean data is clearly a result of pooling of the data from individuals reaching ceiling performance in recovery at different times. Analysis of the trials in which T1 was incorrectly identified yielded an attentional blink for T2 with delayed recovery and a manifest lag-1 sparing effect. This suggests, in the framework of the two stage model, that the duration of gate opening for target recognition requires sufficient certainty of recognition prior to closure, rather than being a fixed, process-related time.