The primary datum was the percent correct for each two-letter comparison (reaction times turned out to be uninformative). For each letter pair there were two percents -- one with each member of the pair as "target" It is important to note that these percents were collected during acquisition. That is, after four days the birds were still far from asymptote on many of the pairs. This tended to spread out the percentage scores, which averaged just above chance (33%) for the most similar letters and above 90% for the most different letters. The experiment thus says something about recognition during formation of the "letter concept": this could be different than, say, similarity judgements involving fully formed concepts.
Each bird yielded a data matrix showing percent correct for each letter paired with all other letters. These matrices were folded, combining the two measures for each letter pair, and averaged. The resulting mass of numbers was rendered meaningful by feeding it to two scaling programs (cf. Shepard, 1980). One of these, ALSCAL, placed the letters into a multidimensional space where similarity is represented by relative proximity. The two-dimensional result appears in Figure 16.1.
Figure 16.1 seems to make some sense, though "dimensions" are not