We began the MI calculation by aligning the first 10 drinking, the first 10 non-drinking, and 20 null trials relative to the cue onset (drinking and non-drinking trials) or a randomly chosen time point during the ITI (null trials). Null trials were constructed from periods of the neural recording that were randomly selected from the ITI such that full null trials did not overlap real trials at any time. At a given time bin t relative to the stimulus onset time and for a given neuron i, we constructed a joint discrete probability distribution pi,t(x,y), where x was the discretized smoothed spiking rate of the neuron (see above, 100-ms bins relative to stimulus onset) and y was either the trial type (real vs null) or the drink outcome (drinking trials vs non-drinking trials). The smoothed spiking rate of the neuron was discretized such that the values across trials were ranked and binned into three states (low, medium, or high firing) with equal number of counts (or as close to equal as possible in the event of tied values). The probability
