Human performance on a choice-reaction time task (Eriksen task) has been simulated by a neural network. In simulations, the network captures many features of normal performance. In addition, changing gain in different layers produces changes that simulate different drug-induced changes. Data from a similar choice-reaction time task have been reanalyzed to test some of the predictions derived from changing gain in different layers. Clonidine antagonizes norepinephrine and acetylcholine activities and changes speed-accuracy tradeoff (i.e., increased frequency of errors at any specified reaction time). That is predicted when gain is reduced in lower layers (attention layer and input layer) of the network. By contrast, manipulating dopamine activity (with pimozide and amphetamine) changes reaction time without changing speed-accuracy tradeoff functions. That is predicted when gain is changed in the output layer of the network.