The previous post was fortunately wrong. The data transform wasn’t correct. That’s what you get for using a powerful data transformation engine without validating the transformations.. Anyway, the nBack error biases are looking much better..
Here’s what I did, I transformed the individual nback responses, clustering temporally by thirds. Now I’m only concerned with the errors. There are two ways to look at this data: # of errors / total # of responses and # of errors / total # of errors. The first gives us the accuracy penalty that each third yields, the second gives us the distribution of errors.
First up: accuracy penalty
Position (p<0.015), When (p<0.04) and the interaction (p<0.03) are all sig.
First of, as usual, the most greatest penalty occurs after rotation (obviously). Most interestingly, when they do rotate, the vast majority of their penalty comes from the tail end of the rotation.
The distribution of errors reflects the same general gist:
The take home message here is that performance is worst for the last 4 responses of the nback during rotation. If we think about this in terms of the pointing task that is going on, it gives us good evidence that subjects are engaging in offline updating of the targets. Specifically, as they come to the end of the rotation they start to figure out where the targets are relative to their current position. This additional processing is killing their nback performance.
So, it’s looking like the offline updating hypothesis is gaining ground. However, if they are sacrificing nback performance in order to determine the target locations, should the rotate pointing time be faster for the WM conditions relative to control? Apparently, this sacrifice of nback performance to track locations isn’t actually gaining subjects much of anything.
