Peak shift revisited: A test of alternative interpretations

Copyright 1991 by the American Psychological Association, Inc. 0097-7403/91/S3.00 Journal of Experimental Psychology: Animal Behavior Processes 1991, Vol. 17, No. 2, 130-i40 Peak Shift Revisited: A Test of Alternative Interpretations This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. David R. Thomas, Kelly Mood, Spencer Morrison, and Eric Wiertelak University of Colorado, Boulder In Experiment 1, 2 groups of human subjects were trained to respond to 1 of 2 light intensity stimuli, S2 or S4, and then were tested for generalization with a randomized series of increasing values from SI to S l l . Both groups, including the group trained to respond to the dimmer value, showed peak shifts to a brighter more centrally located test stimulus. In Experiment 2, which used line angle stimuli, both the size of the difference between S+ and S- and the range of test stimuli that extended beyond S+ were varied. The larger the S+-S— separation and the larger the range, the greater was the peak shift obtained. In Experiment 3, training involved an S— (line angle) surrounded by 2 S+ values with testing symmetrical about the training values and covering either a narrow or a wide range. The wide range produced greater peak shifts in both directions from S-. All 3 experiments support an adaptation-level interpretation of intradimensional discrimination learning and generalization test performance in human subjects. Related work with animals suggests the presence of similar processes. Historically, the central question concerning research and theorizing about (intradimensional) discrimination learning involving stimuli that lie on a single continuum (e.g., size, brightness) has been whether subjects acquire response tendencies to specific stimulus values or learn rules that take into consideration the relational properties of the stimuli. The "absolute" approach was articulated by Spence (1937) who postulated the formation of an excitatory generalization gradient centered about the reinforced stimulus, S+, and an inhibitory generalization gradient centered about the extinguished stimulus, S—. The response tendency associated with any given stimulus is then based on the summation of the generalized excitatory and inhibitory response tendencies at the value of the stimulus in question. Spence's (1937) theory was originally proposed to account for transposition following intradimensional discrimination learning in animals, but it has proved useful in guiding discrimination learning and stimulus generalization research in children and in adult humans as well. A strong prediction from Spence's gradient interaction theory is a peak shift in postdiscrimination generalization gradients (i.e., maximal responding displaced from S+ to a stimulus value further removed from S-). After Hanson (1959) and Thomas (1962) reported peak shifts in work with pigeons, many experimenters sought and found evidence of the same phenomenon in human subjects. Nicholson and Gray (1971, 1972) reported peak shifts (with school children and the line tilt continuum), as did Baron (1973) with adults and tone frequencies and MacKinnon (1972) with adults and lifted weights. As will be pointed out, there are several alternative interpretations of peak shift. A more demanding test of the applicability of Spence's (1937) theory to humans is the nature of the relationship between the amount of separation between S+ and S- and the amount of peak shift obtained. The gradient interaction model predicts a negative relationship, which was found by Hanson (1959) and Thomas (1962) with pigeons and wavelength stimuli. With human subjects, however, the results of such studies have been inconsistent. Baron (1973) using tone frequencies found the predicted negative relationship, whereas Thomas, Svinicki, and Vogt (1973) using light intensity found a positive relationship. The traditional alternative to "absolute" approaches such as Spence's (1937) approach has been the relational view, as espoused by Kohler (1939) and Krechevsky (1938). According to this view, the subject solves the discrimination problem relationally by learning to respond to the larger (or smaller), the brighter (or dimmer), and so on, of the training stimuli. A problem for this simple relational view is that a postdiscrimination generalization gradient shows a peak of response strength (i.e., a single maximal value, beyond which responding decreases as stimuli still farther from S+ are tested). Both the peaked form of the postdiscrimination gradient and the displacement of its peak are consistent with Spence's gradient interaction theory. Although these phenomena are clearly inconsistent with the simple relational view, they are not necessarily incompatible with other relational interpretations of discrimination learning. An alternative relational account of intradimensional discrimination learning was offered by Thomas and his associates (see Thomas, 1974; Thomas et al., 1973) who interpreted peak shift (in experiments with human subjects) using principles based on Helson's (1964) theory of adaptation level. Adaptation level may be thought of as a frame of reference based on an average of all relevant stimuli experienced by the subject. In judgment tasks, all stimuli are thought to be encoded in terms of their relationship to the adaptation level. Discrimination learning procedures and generalization testing procedures may be thought of as judgment tasks. The adaptation-level interpretation was originally developed to explain the "central tendency effect" (i.e., a peak shift toward the center of an asymmetrical generalization test series, which is one with more test values or a wider range of test values to one side of the training value than the other, following training with a single stimulus [see Helson & Avant, 1967; Thomas & Jones, 1962]). It was assumed that the Correspondence concerning this article should be addressed to David R. Thomas, Department of Psychology, University of Colorado, Campus Box 345, Boulder, Colorado 80309-0345. 130 131 This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. PEAK SHIFT REVISITED training stimulus constituted the initial or training adaptation level and that instructions to respond to this value were interpreted as "respond to your adaptation level." During testing with stimuli asymmetrically spaced around the training value, the adaptation level would shift toward the central (average) value in the test series and the peak of responding would shift accordingly. In the case of two-stimulus discrimination training, adaptation level is presumably established approximately midway between the training stimuli, S— and S+, and the subject learns to respond to a stimulus (S+) which is, say, X units greater than the adaptation level. In generalization testing, adaptation level moves toward the central value in the test series, which is normally the previous S+. The subject is presumed to respond during testing in accordance with the "rule" learned during training (i.e., "respond to adaptation level + X"), thus the stimulus value associated with maximal responding during testing is X units greater than the new (prevailing) adaptation level, based on the training plus the testing experience. This tendency results in the peak shift. Thomas (1974) has suggested that the perceptual (adaptation-level) account may be more appropriate than the learning account in experiments with humans, especially when responding is based on instructions rather than reinforcement and nonreinforcement. The adaptation-level account of peak shift is not, however, the only perceptual interpretation available. The generalization gradient may be interpreted through an application of statistical decision theory (see Boneau & Cole, 1967). According to this approach, any given physical stimulus produces a subjective experience (within the organism) called a discriminal process. Because the discriminal process is variable, repeated presentations of the same physical stimulus generate a discriminal distribution. The subject must then establish criteria that define a range of the discriminal distribution that the subject will attribute to a particular stimulus. Blough (1969) has shown that peaked generalization gradients would be found if instructions (for human subjects) or single-stimulus training (for animal subjects) caused them to create two symmetrical response criteria around the mean discriminal process produced by the training stimulus. Blough also noted that peak shift and asymmetrical gradients could be accounted for by assuming that the two criteria were independent. Specifically, a peak shift such as predicted by Spence's (1937) theory would result if discrimination training caused the criterion on the S- side of S+ to be moved toward S+ and had little effect on the opposite criterion. This would mean that stimuli just to the opposite side of S+ from Swould be more likely to produce a discriminal process that would be interpreted as S+ than those on the S— side or the S+ value itself. Although the decision theory account of peak shift is perhaps more appealing than Spence's (1937) theory in dealing with research with adult human subjects, the two views are similar in that both are "absolute" interpretations. They make similar predictions because neither provides a role for the location of the two training stimuli within the range of generalization test stimuli. This factor is critical for the adaptation-level interpretation because this view gives central importance to the relational properties of the stimuli used in training and in testing. The present set of experiments were conducted to test alternative interpretations of postdiscrimination peak shift with adult human subjects. Although these experiments use human subjects, as does most of the literature relevant to adaptation-level theory, principles derived from this theory and from other accounts of psychophysical context effects (see, for example, Parducci, 1974) are clearly also applicable to animals. For example, Giurintano (1972, reported in Thomas, 1974) found evidence of a central tendency shift with line tilt stimuli in pigeons and recently Zoeke, Sarris, and Hofer (1988) using chicken subjects found shifts in choice behavior resulting from generalization testing with stimuli asymmetrically spaced around the training values, a finding clearly consistent with the adaptation-level view. Thus, both Spence's (1937) gradient interaction theory and adaptationlevel theory have been applied, although not always successfully, to both human and infrahuman subjects. The present study is part of a program of comparative research designed to elucidate the boundary conditions for fruitful application of these and other theoretical perspectives. Experiment 1 described in this article demonstrated a peak shift in a direction opposite to that predicted by Spence's (1937) theory. Experiment 2 showed another effect opposite to prediction based on Spence's theory, that of a greater peak shift following training with a more widely spaced S+ and S-. It also showed an effect that is not considered by Spence's theory, a greater peak shift due to an increased range of test stimuli on the side of S+ opposite the S-. Experiment 3 elaborated on this finding, demonstrating that the double peak shift, following discrimination training with an S- surrounded by two S+s, can be interpreted as a "range effect." It had been suggested by Galizio and Baron (1979) that this effect was uniquely consistent with the gradient interaction interpretation. Experiment 1 In Experiment 1, the test series used was asymmetrical with regard to the location of the S+ and S—. This provided a useful test case because it enabled us to create a situation for which Spence's (1937) theory (and its decision theory counterpart) and adaptation-level theory made not just different but opposite predictions. Consider a generalization test series of 11 equally spaced increasing light intensity values, labeled stimulus (S) 1,2,3,..., 11. A group of subjects is instructed (trained) to respond to S4 and not to respond to S2 and then it is tested with the aforementioned series. Spence's theory predicts a peak shifted to a higher intensity value as a consequence of the summation of an inhibitory gradient centered at S2 and an excitatory gradient centered at S4. Decision theory predicts the same shift because of asymmetrical decision criteria (i.e., the lower criterion closer to S4 than the higher criterion). The adaptation-level theory predicts a peak shift in the same direction for yet another reason. According to this theory, subjects have learned to respond to adaptation level (say, S3) plus one unit. As adaptation level shifts toward the center of the asymmetrical test series the peak should shift accordingly. Unlike the alternative theories, however, adaptation-level theory predicts that the shift will develop progressively during the course of testing as the momentary, or 132 THOMAS, MOOD, MORRISON, AND WIERTELAK "prevailing adaptation level," on which the peak shift is based, gradually changes with continuing exposure to the series of mostly higher intensity values. Contrast this situation with that of a group instructed (trained) to respond to S2 and not to respond to S4 and then tested with the same asymmetrical series of stimuli. Both Spence's (1937) theory and decision theory predict a shift toward dimmer rather than brighter stimulus values. The prediction of adaptation-level theory, however, is exactly the opposite. Presumably these subjects have learned to respond to adaptation level minus one unit. If during generalization testing adaptation level shifts to, say, S6, the center of the test This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. series, then the peak of responding should fall at S5 (i.e., it should shift beyond S— to a value further removed from S+). This shift should also develop gradually during testing. These predictions were tested in the following experiment. Method Subjects. The subjecls were 40 students enrolled in introductory psychology courses at the University of Colorado. Apparatus. Each subject was seated 60 cm in front of a 60-cm (wide) x 60-cm (high) panel that was covered with black felt cloth. The experimenter was seated behind the panel and was not visible to the subjects. At approximately eye level, a 2.7-cm (diameter) aperture was present. The subjects viewed a disk of white light projected onto a translucent glass screen behind the aperture. The light source consisted of a light discrimination apparatus manufactured by the Lafayette Instrument Company (Lafayette, Indiana, Model 14011), using a 60-W Sylvania clear Decor Lite (60CA9C/BL). Eleven different light intensities were selected so as to be. 188 log units apart. The intensity values and their experimental designations were SI, 0.83 footlambert (fL, 2.84 cd/m2); S2, 1.28 fL (4.39 cd/m2); S3, 1.97 fL (6.75 cd/m2); S4, 3.04 fL (10.42 cd/m2); S5, 4.69 fL (16.07 cd/m2); S6, 7.23 fL (24.77 cd/m2); S7, 11.2 fL (38.38 cd/m2); S8, 17.2 fL (58.93 cd/m2); S9, 26.5 fL (90.80 cd/m2); S10, 40.9 fL (140.14 cd/ m2); and SI 1,63.0 fL (215.86 cd/m!). The experiment was conducted in a small, dimly illuminated room. The light reflected from the disk when it was not illuminated was approximately .01 fL. Procedure. Approximately two thirds of the subjects in this experiment were women. The subjects were unsystematically assigned to two groups (« = 20) with a similar distribution of the sexes in each. After the subject was seated in front of the stimulus panel, the following instructions were read: "This is an experiment in brightness perception. A light will be presented repeatedly through a small hole in the screen in front of you. Each time it will be presented for 3 s and may have a different brightness. The first brightness is called the test brightness. Try to remember this brightness because you will have to distinguish it from all the other brightnesses. When you do recognize the test brightness, press the response button. If a subsequent brightness is different from the test brightness, do not press the button. Remember, each time the light is presented it will stay on for only 3 s, so try to respond while the light is on. Indeed, try to respond as quickly yet as accurately as you can. I will tell you whether you are correct on the first few trials; then you will continue without further help. The first light is the test brightness. Keep its brightness in mind. For every light after that, respond by pressing the button as quickly as you can if and only if it is the same as the original light. Any questions?" Only questions dealing with the procedure were answered by the experimenter. Following any needed clarifications of instructions, each subject in both groups was shown the appropriate S+ for their group. On the next training trial, the S- value was presented. For both groups, training continued for 24 trials with S2 and S4 presented 12 times each in unsystematic order. After each correct response, the experimenter said "correct," and after each incorrect response, the experimenter said either "no, that was different from the original" or "no, that was the same as the original," depending on the error. Next, with no interruption in the procedure, the subjects were shown six series of all 11 stimuli; feedback was no longer given. The stimuli were randomized within each series in both discrimination training and generalization testing, and the interstimulus interval ranged unsystematically from 2 to 3 s. Two different random sequences were used in generalization testing with half of the subjects in each group randomly assigned to each sequence. If the subject made more than two errors in the last eight training trials, an additional eight training trials were given. Subjects who made more than two errors during these additional trials were eliminated from the experiment and replaced. There were 3 such subjects in the S4+ S2- group and 4 in the S2+ S4- group. Four of the subjects in each group required the extra training trials to meet the performance criterion. There were no systematic differences in generalization test performance between subjects that required the extra training and those that did not, and no systematic differences in test performance related to differences in training performance among subjects who met the performance criterion during initial training. Results and Discussion For subjects in each group, the mean number of responses to each of the 11 test stimuli was calculated and the result is presented in Figure 1. Consider first the gradient of the group trained with S2 as S— and S4 as S+. As expected, this group showed a peak shift from the S+ to a greater intensity value with responding decreasing to the most intense stimuli. This result, of course, is consistent with all three accounts of the peak shift. Recall, however, that only the adaptation-level account predicts that the shift should be progressive during the course of testing as continued exposure to the test values results in continual recalculation of the adaptation level. Although the shift in adaptation level may occur early in testing it may be possible to observe the progression of the shift by comparing responding in the first test series with that in the last test series. Because each stimulus is presented once within each test series, the gradient indicates the number of subjects who responded in the presence of each test stimulus. Table 1 presents the generalization gradients of the S4+ S2- group based on the first test series and the last (sixth) test series. In accordance with the adaptation-level prediction, an increasing tendency to respond to higher intensity values during testing may be seen. The shift is most apparent in the reduction of responding to the positive training value, S4, from 11 out of 20 subjects in the first series to 4 in the last series, (x2 = 5.23, p < .05) and the corresponding increases in responding to S8 and S9 from 4 and 5 subjects, respectively, to Hand 12. (For S8, x2 = 10.10,p< .01; for S9, x2 = 5.01, p<.05.) Because the adaptation level presumably shifts within the first test series, it may be possible to demonstrate this effect as well. Toward this end, two different sets of random stimulus presentations were used in this experiment with 10 subjects assigned to each set. In one set, S7 was the second test stimulus experienced; the first was the training value, S4. In the second PEAK SHIFT REVISITED •—• 2+4- This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. O—0 4+2- 1 2 3 * 5 6 7 8 9 1 01 1 STIMULUS VALUE Figure 1. Mean light intensity generalization gradients (i.e., responses to the different test stimuli by subjects in Group 1, S4+ S2-, and Group 2, S2+ S4-) of Experiment 1. set, S7 was the eighth stimulus experienced. When S7 was second, only 3 subjects responded to it, whereas when it was eighth, and the subjects had experienced four test values of even greater intensity, all 10 subjects responded to it. This difference is significant: \2 = 10-77, P < .01. Thus, there is no doubt that the peak shift in the S4+ S2- group is not fixed at the end of training but increases during the course of testing. Now, consider the generalization gradient of the subjects in the S2+ S4- group, also presented in Figure 1. In accordance with the adaptation-level account, the peak of this gradient falls at S5 (i.e., it is shifted from the S+ [S2] to a value beyond the S- [S4] in the direction opposite that predicted by Spence's [1937] gradient interaction theory). As was the case for the S4+ S2— group, it is possible to trace the progression of the peak shift in the S2+ S4- group. Table 1 presents the Table 1 Generalization Gradients* Stimulus value Series 1 2 1 6 0 0 0 0 1 1 1 6 7 6 7 9 10 9 3 4 5 6 7 8 9 10 11 4 14 5 12 2 3 0 2 1 5 1 4 0 1 0 0 Group S 4-1- S 2- 11 4 13 10 16 12 13 13 Group S 2+ S 4- 9 12 18 14 13 13 2 7 * Number of subjects responding to the different test stimuli during the first and sixth test series of Experiment 1. 133 gradients of this group based on the first and the last (sixth) test series. The most dramatic changes are seen in the tendency to respond to the most intense stimuli, S7 and above. Whereas in the first test series a total of four responses was made to these stimuli, in the sixth series 17 responses were made to these stimuli. None of the changes in responding to these individual stimuli (S7-S10) was significant, however, with the largest x2 (for S7) = 3.58, p > .05. With maximal responding already at S5, clearly much of the shift in responding had already taken place during the first test series. Indeed the evidence of a shift within the first test series was quite dramatic. When S4 (i.e., the S- in training) was the first test stimulus experienced, only 1 of 10 subjects responded to it. When it was seventh and the subjects had seen four stimuli of even greater intensity, 7 of the 10 subjects responded to it: X 2 = 7.50,p<.01. Because the gradients of both the S2+ S4- and the SV4+ SV2- groups showed peak shifts toward more intense stimulus values, it is appropriate to consider the possibility of an energizing effect of stimulus intensity as proposed in Hull's (1949) concept of stimulus intensity dynamism. There is nothing in the dynamism concept that suggests that the shift would be progressive during testing. More to the point, Thomas et al. (1973) showed that there is a peak shift to a less intense stimulus when S+ was more intense than S-, that most test stimuli were less intense than both training values, and that there is no peak shift at all in a group given single stimulus training and tested with intensity stimuli symmetrically spaced around the S+. Thus, there is no evidence of a dynamism effect in this experimental situation. The results of the present experiment provide strong support for the adaptation-level account of peak shift following intradimensional go-no-go discrimination training with human subjects. The application of gradient-interaction theory and statistical-decision theory fall short because they give no role to the frame of reference provided by the generalization test series and, in particular, to the location of the S+ and S— values within that series. Experiment 2 In Experiment 1 the location of S+ and S— were reversed in the two groups but consisted of the same two stimulus values, S2 and S4. Thus, the adaptation level based on training was presumably the same in both groups, and the shift in adaptation level during testing was presumably also the same. The shift in adaptation level from its training value to its value during testing presumably determined the size and the direction of the peak shift. If this shift were larger, the displacement of the peak should be greater. This effect has been reported in an experiment by Thomas, Strub, and Dickson (1974) in which subjects were exposed to a single (dim) intensity value and then were tested for their ability to recognize this value with increasingly asymmetrical test series (in different groups). Whereas the training stimulus was SI in all cases, for one group the test values were SI, S2, and S3; for another group, they were SI, S2, S3, S4, and S5; and so on. Testing with an asymmetrical series produced a central tendency shift, and in accordance with adaptation-level theory, This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. 134 THOMAS, MOOD, MORRISON, AND WIERTELAK the greater the asymmetry, the larger the shift that was observed. Furthermore, for different groups in this experiment, the adaptation level of the different test series was determined by the stimulus rating method (see Helson, 1964, chap. 4), and group differences in the location of the peak of responding were directly related to group differences in measured adaptation level. If the peak shift following intradimensional go-no-go discrimination training is due to a change in adaptation level from its training to its testing value, then it should be possible to modulate the size of the peak shift by varying the degree of asymmetry in the test series used. Evidence that suggests that this is so comes from a comparison of the results of Experiment I with those of an experiment previously reported by Newlin, Rodgers, and Thomas (1979, Experiment 1), which was performed in this laboratory with the same apparatus, procedure, physical stimulus values, and so on. In that experiment, a group of subjects was trained to respond to S2 and not to respond to S4 and then they were tested with the range S1-S9. The peak of this group's gradient fell at S4, the S— in training, thus the peak shift was in the same direction (opposite that predicted by Spence's [1937] theory) but not as large as in the present Experiment 1 in which the test series was more asymmetrical (range S1-S11). Because there were differences in the amount of training administered in the two different experiments, as well as other minor procedural differences, it is appropriate to repeat this comparison within a single experiment and to do so with a greater difference in the degree of asymmetry in the generalization test series. This was one of the purposes of Experiment 2. If responding during testing is based on application of the rule, "respond to adaptation level + X," then the amount of peak shift should be governed both by the amount of change in adaptation level and by the size of X. As we have indicated above, and as was demonstrated empirically by Thomas et al. (1974), the amount of change in adaptation level can be manipulated by varying the degree of asymmetry in the generalization test series. The size of X can be manipulated by varying the distance between the S+ and S- used in training. The larger the separation, the farther each stimulus is from the adaptation level that is approximately midway between them, thus the larger is X. Clearly, then, adaptationlevel theory predicts that the more widely separated the S+ and S— the greater will be the peak shift. This prediction is opposite that made by Spence's (1937) theory and by decision theory. Furthermore, experiments with pigeon subjects by Hanson (1959) and by Thomas (1962) have unequivocally supported Spence's position. In contrast to this finding with animal subjects, evidence for a positive relationship between S+-S— separation and amount of peak shift was obtained in an experiment with human subjects reported by Thomas et al. (1973), using different intensities of white light as stimuli. When the S+ was more intense than the S—, the results clearly showed increasing peak shift with increasing S+-S— separation. When the S+ was less intense than the S—, the results were less clear, with only the largest S+-S- separation producing a measurable shift to a value less intense than S+. Because of the discrepancy between the findings with animal versus human subjects and in view of the opposite predictions made by the different theories, it is appropriate to reexamine the relationship between S+-S— separation and amount of peak shift. To increase generality, we decided to do so with a qualitative rather than a quantitative stimulus dimension. Thomas and Thomas (1974) had reported central tendency shifts with the dimension of angular orientation of a line viewed on a tachistoscope screen, and we adapted this experimental situation for the present experiment. In Experiment 2, we examined the two proposed causes of peak shift using line angle stimuli. Two groups were trained with widely spaced S+ and S— stimuli (S4 and SI, respectively) and were tested with a narrow range (S1-S7) and a wide range (SlSl 1) of test stimuli, respectively. A third group was also tested with the wide range but their training stimuli were more closely spaced (i.e., S4 was still S+ but S— was S2). It was predicted that Group 1 (narrow range) would show less of a peak shift than Group 2 (wide range) and that Group 3 (smaller S+-S— separation) would show less of a peak shift than Group 2 (wider S+-S— separation) No comparison is made of the gradients of Groups 1 and 3, because with differences in both training and testing stimuli any obtained differences in generalization performance would be uninterpretable. Method Subjects. The subjects were 60 students enrolled in Introductory Psychology courses at the University of Colorado. Apparatus. The apparatus was a Scientific Prototype (New York) two-field tachistoscope (Model 800F). A screen separated the experimenter from the subjects so that they could not see the experimenter or the stimuli before they appeared on the tachistoscope screen. Procedure. The subjects were unsystematically assigned to three groups (n = 20), with a similar distribution of the sexes in each. Training was the same for subjects in Groups I and 2; only the test procedure was different. For these subjects the S+ (S4) was a black line 3.81 cm x .32 cm oriented at 55° counterclockwise rotation from horizontal and centered in a circular field 7.62 cm in diameter. The S- (SI) was a 70° line. The lines were presented with a 1-s exposure time and the intertrial interval was 8-12 s. The instructions were modified appropriately from those used in Experiment 1, but the training procedure was otherwise identical. So was the test procedure, with the exception of the number and values of test stimuli used. For the "narrow range" group (i.e., Group I) test values of 70°, 65°, 60°, 55°, 50°, 45°, and 40° were arranged in nine different random series. For the "wide range" group (i.e., Group 2), in addition to those values, 35°, 30°, 25°, and 20° stimuli were used. The 11 different stimuli were arranged in six different random series. Thus, the total number of test stimuli was similar but not identical for these two test groups. Group 3 was trained with S4 (55°) as S+ but S2 (65°) as S-. These subjects received the same test as Group 2, experiencing the wide range of test stimuli (i.e., SI-SI 1). As in Experiment 1, if the subject made more than two errors in the last eight training trials, an additional eight training trials were given. Subjects who made more than two errors during these additional trials were eliminated from the experiment and replaced. There were 3 such subjects in each of Groups 1 and 2 and 5 such subjects in Group 3. There were also 4 subjects who required extra training 135 PEAK SHIFT REVISITED in Group 1, 3 in Group 2, and 4 in Group 3. There were no systematic differences in generalization test performance as a function of how well subjects performed in training. This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. Results and Discussion Because of between-group differences in the number of opportunities given to respond to the different test stimuli, the data were analyzed in terms of the probability of responding to each stimulus rather than the absolute number of responses. Consider first the gradients of Groups 1 and 2, as shown in Figure 2. According to Spence's (1937) gradient-interaction theory and statistical-decision theory, the peak shift is determined by training parameters, such as the separation between S+ and S—. Yet both of these groups were trained identically, with the same S+ and S-, the same amount of training, and so on. The figure reveals, however, in agreement with prediction based on adaptation-level theory, that the wide range group (Group 2) shows a larger peak shift than the small range group (Group 1). Indeed the gradient of Group 1 shows no peak shift at all but only asymmetry with more responding to higher than to lower stimulus values. The obtained difference in amount of peak shift in the two groups replicates, with a single experiment, findings based on a comparison of different experiments as described earlier. Although the predictions of the different theories have been presented in terms of the location of the peak, it should be realized that a shift of the peak reflects a shift in the entire distribution of responses. Thus, to properly evaluate the results, a measure is needed that is sensitive to changes in the entire distribution of responses along the stimulus continuum and not just in the mode. Such a measure may be obtained by treating each individual subject's gradient as a grouped S+ 1 2 3 4 5 • • GROUP 1 (NARROW RANGE) • • GROUP 2 (WIDE RANGE) 6 7 8 9 1 01 1 + STIMULUS VALUE Figure 2. Mean line angle generalization gradients (i.e., responses to the different test stimuli by subjects in Group 1 [narrow range) and Group 2 [wide range]) of Experiment 2. frequency distribution and computing its mean; the higher that mean, the greater the shift in responding toward higher stimulus values. This measure was initially used in a peak shift study with pigeons by Thomas (1962) and has been widely used since then because it provides a sensitive and continuous measure of changes in the locus of responding along the test continuum. In the computation of the individual gradient means of Groups 1 and 2 only responding to SlS7 was included because both groups were exposed to these stimulus values. The group mean of the individual gradient means of Group 2 (wide range) computed in this manner was 5.45 compared with 4.19 for Group 1 (narrow range), and this difference was highly significant, £(38) = 8.45, p < .01. The comparison of the gradients of Groups 2 and 3 provided a test of the hypothesis that the size of the peak shift varies positively with the amount of S+-S— separation in training. As is shown in Figure 3, in agreement with the results of the Thomas et al. (1973) experiment, which used light intensity stimuli, the present experiment also revealed a positive relationship, with Group 2, with S4 as S+ and S1 as S—, showing a larger shift than Group 3, with S4 as S+ and S2 as S-. To statistically compare the gradients of Groups 2 and 3, the gradient means were again used. This time the entire gradients of subjects in Group 2 based on responding to all 11 stimulus values were used. The mean of the gradient means was 5.98 for Group 2 compared with 5.03 for Group 3, and this difference was highly significant, ((38) = 4.34, p < .01. Thus, the positive relationship between S+-S— separation and amount of peak shift is not unique to the use of a quantitative stimulus dimension (i.e., light intensity) but exists also with line angle stimuli. Experiment 3 The evidence in support of the adaptation-level interpretation of peak shift in human subjects is so strong that it is appropriate to reconsider the evidence against it. Using light intensity stimuli, White and Thomas (1979) obtained evidence of a double peak shift in a task in which a single Swas surrounded by two S+ values. In this situation the obtained generalization gradient is bimodal, but each of the peaks is displaced from the nearby S+ to a stimulus value farther removed from the S-. The generalization test values are symmetrical around the training stimuli, thus there should be no change in adaptation level from training to test. If a change in adaptation level from training to test is the only basis for a peak shift, then none should occur in this situation. Even if adaptation level were to increase or decrease during testing, it could not account for a peak shift in both directions. Galizio and Baron (1979) pointed out that the double peakshift phenomenon is inconsistent with the adaptation-level formulation for the reasons indicated, whereas it is predictable from Spence's (1937) gradient-interaction model. Perhaps, however, with some modification, the adaptation-level position might yet prove a useful alternative. The notion that subjects learn and subsequently apply the adaptation level + X rule (see Thomas et al., 1973), or in the three-stimulus case adaptation level ± X, carried with it the 136 THOMAS, MOOD, MORRISON, AND WIERTELAK >t tt o. Ill 0. m ^^x 0.75 c^ / / / / ( i t / I — ' v1V\ x * V \ » ^x ^ .[' — \ ^ \ \\ \w TI \ v U i i i i i i j i z This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. * y < / 0.50 UI 2 o GROUPS 4+2- — (A Z o • GROUP 2 4+ 1- o S+ 1.00 5 m o • 0.25 \ Q \ \ \ \ / r D'/ S 1 \ • \ \ \ S / x -* £'-f 1 1 3 2 4 i i i i 5 6 7 8 \ V v ^ r^-x 9 ^ 1 01 1 STIMULUS VALUE Figure 3. Mean line angle generalization gradients (i.e., responses to the different test stimuli by subjects in Group 2, S4+ S1 -, and Group 3, S4+ S2-) of Experiment 2. assumption that X remained constant during generalization testing. Perhaps it is this assumption that needs to be modified. The supposition that X remains constant is inconsistent with the basic tenet of adaptation-level theory that stimuli are always judged relationally. Suppose the subject were trained, as in Experiment 1, to respond to S4 and not to S2. What the subject learns is to respond to a value that is slightly brighter than adaptation level. But what constitutes slightly brighter? If the subject sees much brighter stimuli during the test, he or she might alter the concept of slightly brighter so that it applies to more intense stimuli than it would if he or she were not exposed to those more extreme values. In other words, X may increase in proportion to increases in the range of test stimuli used. This assumption is consistent with the results obtained in Experiment 2. Either a change in adaptation level, in X, or in both could account for the difference between the gradients of Group 1 and Group 2. Note that there already exists in the human research literature evidence of the effect of the range of test stimuli on the level of generalized responding. Thomas and Hiss (1963) trained their subjects to respond to a particular wavelength of light (530 nm) and then tested different groups for generalization over three different ranges, all symmetrical with regard to the training value. They found that the tendency to respond to a particular test stimulus, say 550 nm, was greater for subjects for whom the generalization test included stimuli farther removed from S+ (e.g., 570 nm) than for subjects for whom 550 nm was the most extreme value used in testing. Parducci (1974) has proposed a theory to account for the effect of the range of stimuli and the frequency of their presentation in judgment tasks and has observed such effects in a wide variety of circumstances. Thus, the proposal that X in the adaptation level + X formulation might vary in proportion to the range of test stimuli is certainly tenable. In view of the overwhelming evidence that changes in adaptation level can produce peak shifts, it is necessary to establish that changes in X may also occur. Because only changes in X could account for the double peak shift observed in the three-stimulus situation, that paradigm permits a test of this hypothesis. We predicted that by expanding the range of test stimuli symmetrically spaced around the training values, we could increase the displacement of the two peaks from the training S+ values. If this hypothesis were supported the presence of the double peak shift would no longer stand as an embarrassment to the adaptation-level position. Two considerations led us to use line angle rather than light intensity in designing this experiment. First, there was the asymmetry in the results obtained with light intensities by Thomas et al. (1973). Recall that they found that it was much easier to produce peak shifts toward brighter than toward dimmer values. Second, the use of line angles provided a larger range of discriminable test values than was available with our visual intensity apparatus. New line angle stimuli were prepared that covered the entire range from horizontal (180°) moving clockwise through vertical (90°) to 5° counterclockwise from horizontal. These stimuli were designated SlS36. There was no S37 (i.e., 0°) because that is the same as 180° (i.e., horizontal). As will be seen, not all available line angles were used in the experiment. Method Subjects. The subjects were 40 students enrolled in Introductory Psychology courses at the University of Colorado. Apparatus. The apparatus was the same as that used in Experiment 2. Procedure. The subjects were assigned unsystematically to two groups with a similar distribution of the sexes in each. All subjects received the same training. The instructions were modified to indicate that there were two test values and the subjects were to respond by pressing a button as quickly as possible when they recognized either one. These values were 100° (now designated as SI 7) and 80° (now designated as S21), and they were shown on the first two training trials. In total, 30 training trials were given with 10 presentations each of SI7, S21, and S19 (90°, the S-) in a nonsystematic order. As in the previous experiments, feedback was only given during training. Because the S- was so easy to identify, only one subject made more than three errors during training. This subject was replaced. When training was completed, with no interruption in the procedure generalization, testing was begun. Subjects in Group I (narrow range) were tested with five random series, each of which included one presentation of S13 (120°), S15 (110°), S16 (105°). S17 (S+, 100°), S18 (95°), SI9 (S-, 90°), 520(85°), S21 (S+, 80°), S22 (75°), S23 (70°), and S25 (60°). The stimuli in the region where the peak was anticipated to fall were more closely spaced than at the extremes to maximize sensitivity to differences between the groups. Subjects in Group 2 (wide range) were tested with three different random series that included each of the above values plus S7 (150°), S9 (140°), S l l (130°), S27 (50°), S29 (40°), and S31 (30°). Thus, the total number of test stimuli administered was similar but not identical in the two groups. Results and Discussion As in the previous experiments, there were no systematic differences in test performance as a function of the level of 137 This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. PEAK SHIFT REVISITED training performance. Because of differences between groups in the number of opportunities given to respond to the different test stimuli, the data were analyzed in terms of the probability of responding to each stimulus rather than the absolute number of responses. The mean probability gradients of the two groups are presented in Figure 4. As may be seen, the gradient of Group 1 (narrow range) shows the double peak shift with maximal responding displaced from both training values to the adjacent value farther removed from S-. This replicates the finding of a double peak shift with light intensity stimuli, as reported by White and Thomas (1979) and extends that finding to a new (qualitative) dimension, line angle. Note that the gradient of Group 2 (wide range) shows more extensive peak shifts than that of Group 1 in both directions. Because the two groups were trained identically, the difference in the amount of displacement of the two peaks can only be attributed to a change in the value of X from training to test. To statistically analyze the differences between the two groups' gradients, a procedure was used that is an elaboration of the one used in Experiment 2. In this case, two means were obtained from each gradient, one based on stimuli from S19 through S25 and one based on stimuli from S19 through S13. Because our concern was with how far a given stimulus was from S- (and S+) and not in which direction that stimulus differed from S-, the gradient of each subject was "folded over" with each half (i.e., the left half and the right half) treated as a replication. Thus, responses to S18 were treated as if they had been made to S20 on the second replication, those to S17 were treated as if they had been made to S21, and so on. This procedure enabled us to both directly compare the amount of shift in the two directions and pool data over the two directions in the comparison of amount of shift as a function of narrow versus wide range of test stimuli. As the figure suggests, there is neither any effect of direction from S- (F< 1, for replication) nor any interaction between group and replication (F< I ) . Pooling across replications, the mean of the gradient means was 22.1 for the narrow range group •—• GROUP 1 (NARROW RANGE) •- -• GROUP 2 (WIDE RANGE) CL 0.6 I 15 13 17 + 1921+ 23 27 31 16 18 20 22 25 29 STIMULUS VALUE Figure 4. Mean line angle generalization gradients (i.e., responses to the different test stimuli by subjects in Group 1 [narrow range] and Group 2 [wide range]) of Experiment 3. versus 22.8 for the wide range group, F(\, 38) = 10.05, p < .01. Although it is significant, the size of the difference between the amount of peak shift observed under narrow range and wide range conditions is quite small. Note, however, that although the wide range is twice as great as the narrow range (i.e., 60° vs. 120°), the range experienced in training was only 20°, so both test ranges are far wider than the subjects experienced in training (i.e., three times and six times as wide). For simplicity of exposition, we postulated that the size of X would vary in proportion to the size of the range. All that is necessary to account for the double peak shift is a positive relationship, which is what we observed. General Discussion Experiment 1 showed that a peak shift can be produced that is in the direction opposite to that predicted by Spence's (1937) gradient-interaction theory and by statistical-decision theory as well. Furthermore, the shift developed gradually during the course of asymmetrical generalization testing. This pattern of results can only be accounted for by adaptationlevel theory. Experiment 2 explicitly tested the notion that responding during testing reflects the application of the adaptation level + X rule that the subjects learned during training. The amount of change in adaptation level was manipulated by varying the asymmetry of the test series, whereas the size of X was manipulated by varying the S+-S- separation used in training. Both manipulations had the predicted effect. The effect of narrow versus wide range (i.e., of the degree of asymmetry in the test series) replicates the finding by Thomas et al. (1974) that the size of the central tendency shift, which reflects a change in adaptation level from training to test, varies directly with the size of the range. The positive relationship between S+-S— separation and amount of peak shift replicate"! an earlier finding by Thomas et al. (197 3) who used light intensity stimuli. The earlier finding might have been viewed as somewhat equivocal because of an asymmetry in the results (i.e., larger differences when the S+ was more intense than S- than when the S- was more intense). It is appropriate to point out, therefore, that this asymmetry is entirely consistent with the adaptation-level formulation. To simplify the explication of the adaptation-level position, we have made the assumption that the adaptation level will be located at the center of the generalization test series (or midway between the training values). Adaptation level is the "psychological" center but not necessarily the physical center of the test series, because some stimulus values may be weighted more heavily than others in the computation of "average." Thomas et al. (1973) empirically determined the adaptation level for their generalization test series through the category rating method and found it to be at a value substantially more intense than the central test value. This accounts for the asymmetry in their obtained results. Substantial shifts were obtained when the S+ was brighter than the S-, and both training values were relatively dim because change in adaptation level from training to test was enhanced for these groups. On the other hand, when the S+ was less intense than This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. 138 THOMAS, MOOD, MORRISON, AND WIERTELAK the S-, and both training values were relatively bright, less shift (or no shift) was obtained because the adaptation level of the entire test series was not much lower than was the adaptation level established during training (except for the group with the widest S+-S- separation). It may be necessary for the change in adaptation level from training to test to be rather substantial to produce a peak shift. This may explain why Group 1 in Experiment 2 with the restricted range of test stimuli showed an asymmetrical gradient but not a peak shift. The amount of change in adaptation level from training to test may have been insufficient to produce a peak shift. In addition to results with line angles in the present study and with light intensities by Thomas et al. (1973), a positive relationship between S+-S- separation and amount of peak shift was also reported by Doll and Thomas (1967) with wavelength stimuli, with the exception of one group for which the S— was in a different color name category than the S+. The Baron (1973) study with tone frequencies is the only one of which we are aware that obtained the negative relationship predicted by Spence's (1937) theory. It seems unlikely that the difference in modality is critical, and the Baron experiment should probably be replicated. The claim that X remains constant, whereas only adaptation level may change from training to test was another simplifying assumption that was sufficient to yield predictions that were consistently supported by the accumulating data base (see Thomas, 1974). Only the demonstration of the double peak shift called this assumption into question. The results of Experiment 3 provide an explanation of this phenomenon that is entirely consistent with the main tenet of the adaptation-level position (i.e., that stimuli are judged relatively rather than absolutely). It simply extends this principle one step further to account for frame of reference effects on the judgment of the size of X. Given the demonstration that X may vary with the range of generalization test stimuli, it is important to acknowledge again that the larger peak shift observed with a larger range in Experiment 2 could be due to a larger change in adaptation level, a larger value of X, or both. There is no way to distinguish between these possibilities, so it is important to note that each may have played a role. The effect observed in Experiment 3 could only be due to a change in the value of X, whereas only the change in adaptation level is consistent with the gradual shifting of the peaks observed in Experiment I. Both adaptation-level theory and Parducci's (1974) rangefrequency theory provide interpretations of frame of reference or psychological context effects in judgment tasks. These effects demonstrate that stimuli are judged relationally, not just with regard to each other, but also with regard to other stimuli experienced in the experimental situation. Parducci's theory is clearly applicable when context effects occur with generalization test stimuli that are symmetrically spaced around the training values but that vary in their range or their frequency of presentation. The adaptation-level theory clearly applies with generalization test stimuli that are asymmetrical with regard to the training values. The explanation of the double peak-shift phenomenon requires a combination of both of these approaches. Although the double peak-shift phenome- non is consistent with Spence's (1937) theory the present interpretation is far more parsimonious because it accounts also for the effect of the range of test stimuli, a parameter about which Spence's theory has nothing to say. It is time to return to the original question of whether intradimensional discrimination learning is relational or absolute and the related question of whether the answer depends on the species of subject used. The present set of experiments demonstrates unequivocally that human subjects who learn successive intradimensional discriminations may learn rules that take the relational properties of the training stimuli into account and then apply those rules in subsequent generalization testing. We do not deny, however, that subjects may also learn about the absolute values of the stimuli used in training. Premack (1978) pointed out that the dichotomy between absolute versus relational learning interpretations was inappropriate because learning could occur at both levels simultaneously and which level would be revealed would depend on the test conditions, the species used, and so on. It also depends on the nature of the stimulus dimension used, in particular on the degree to which absolute identification of individual stimulus values is easy or difficult. Our use of the light intensity dimension in most of the experiments that support adaptation-level theory was predicated on the supposition that, because absolute identification of intensity values would be difficult, the use of a relational strategy or rule would be encouraged. The use of line angle stimuli is particularly informative in this regard because subjects can be instructed to encode a particular line angle, say 60°, in an absolute fashion, say by visualizing a clock hand pointing at 1 o'clock, which prevents a central tendency shift that otherwise would occur (see Thomas & Thomas, 1974). Because Spence's (1937) theory was proposed to apply to animal subjects, and adaptation-level and range-frequency theories were developed in work with humans, it is appropriate to ask whether the species difference is critical, as Premack (1978) has suggested. The answer would appear to be both yes and no. Our results lead us to conclude that in work with human subjects, Spence's theory has provided a useful heuristic purpose, suggesting experiments the results of which are better explained by other formulations. The only evidence we have found for absolute rather than relative learning strategies is when the use of language intervenes (i.e., when stimuli can be easily labeled [e.g., a clock hand pointing at 1 o'clock]). This is not the sort of thing that Spence had in mind. It is probably fair to say that most of the animal literature contains results consistent with Spence's (1937) "absolute" approach to discrimination learning. A good example is the studies described earlier that found a negative relationship between S+-S- separation and amount of peak shift. This does not mean, however, that adaptation-level effects cannot also be demonstrated in animals. In an early attempt to find a central tendency shift in pigeons, Thomas and Barker (1964) trained pigeons to respond to a particular wavelength and then tested them with stimuli that were symmetrically spaced around S+ in one group or maximally asymmetrical (i.e., all test values other than S+ were shorter wavelengths) in another group. The obtained gradients were comparable, with no suggestion of a shift in the asymmetrical test group. Thomas 139 This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. PEAK SHIFT REVISITED (1974) subsequently pointed out that the difficulty in demonstrating a central tendency shift in pigeons is, in fact, consistent with the adaptation-level formulation. If pigeons are trained for many hours to respond to a particular stimulus and then they are given an asymmetrical generalization test in extinction that lasts an hour or less, the substantially greater weight of the training experience may make any change in adaptation level during testing too small to be detectable. Giurintano (1972, reported in Thomas, 1974) demonstrated with humans that longer training reduced the size of the central tendency shift obtained, and Giurintano was able to obtain a central tendency shift in pigeons by limiting training with a line angle stimulus to 30 m or less (following extensive training with an orthogonal, wavelength stimulus). One wonders whether the negative relation between S+-S— separation and amount of peak shift found in animal work would still attain if a procedure were developed that enabled pigeons to acquire the discrimination very rapidly followed by extensive asymmetrical generalization testing. The tide is beginning to turn as new techniques are developed that are sensitive to context effects in nonhumans. Several researchers have reported range effects in pigeons using quite different procedures and sensory continua. For example, Chase (1984) examined an analog of the method of absolute judgment (identification) using differing levels of light intensity and Hinson and Lockhead (1986) reported range effects in a go-no-go discrimination task with flicker frequency stimuli. Zoeke, Sards, and their colleagues (see Sards, 1990; Zoeke, Sards, & Hofer, 1990) have observed both adaptation-level and range effects in chickens using a choice procedure with size of a wooden box on the floor of the operant chamber as the stimulus continuum. In one of their experiments, chickens learned to respond at the left key when the stimulus was a small box and at the right key when the stimulus was a larger box. After mastery of this task, the birds were tested with increasingly larger boxes, and they gradually shifted their responding, pecking the left key for box sizes that initially were responded to at the right key. Thus, boxes that were initially categorized as "large" (i.e., appropriate for a response at the right key) were later categorized as "small" when subjects had experienced far larger boxes during testing. In the Zoeke et al. study, unlike the central tendency shift study of Giurintano (1972, reported in Thomas, 1974) all responses were reinforced during generalization testing, which permitted testing over six extended sessions. The gradual shift in choice responding that was observed almost certainly has the same underlying basis as the peak shifts in Experiments 1 and 2 in the present article (i.e., a shift in adaptation level from its training value, between the two training stimuli, to a test value more centrally located in the asymmetrical generalization test series). Although it it easier to demonstrate relative stimulus encoding in humans than in animals, it is now clear that this is not an exclusively human phenomenon. We concur with Premack's (1978) view that discrimination learning may be absolute or relative, although we disagree that the species used is the critical determinant. Clearly further work will be required to specify what the important determinants are. The contribution of the present experiments is to demonstrate the power of frame of reference manipulations in experimental paradigms not typically viewed from a relational perspective. References Baron, A. (1973). Postdiscrimination gradients on a tone continuum. Journal of Experimental Psychology, 101, 337-342. Blough, D. S. (1969). Generalization gradient shape and summation in steady-state test. Journal of the Experimental Analysis of Behavior, 12, 91-104. Boneau, C. A., & Cole, J. L. (1967). Decision theory, the pigeon, and the psychological function. Psychological Review, 74, 123-135. Chase, S. (1984). Pigeons and the magical number seven. In M. Commons, R. J. Herrnstein, & A. R. Wagner (Eds.), Quantitative analysis of'behavior (Vol. 4, pp. 35-57). Cambridge, MA: Ballinger. Doll, T. J., & Thomas, D. R. (1967). Effects of discrimination training on stimulus generalization in human subjects. Journal of Experimental Psychology, 75, 508-512. Galizio, M., & Baron, A. (1979). Human postdiscrimination gradient: The effects of three-stimulus discrimination training. Animal Learning & Behavior, 7, 53-56. Hanson, H. M. (1959). Effects of discrimination training on stimulus generalization. Journal of Experimental Psychology, 58, 321-334. Helson, H. (1964). Adaptation level theory. New York: Harper & Row. Helson, H., & Avant, L. L. (1967). Stimulus generalization as a function of contextual stimuli. Journal of Experimental Psychology, 73, 565-567. Hinson, J. M., & Lockhead, G. R. (1986). Range effects in successive discrimination. Journal of Experimental Psychology: Animal Behavior Processes, 12, 270-276. Hull, C. L. (1949). Stimulus intensity dynamism (V) and stimulus generalization. Psychological Review, 56, 67-76. Kohler, W. (1939). Simple structural functions in the chimpanzee and in the chicken. In W. D. Ellis (Ed.), A source book ofGestalt psychology (pp. 217-227). New York: Harcourt, Brace. Krechevsky, I. (1938). A study of the continuity of the problem solving process. Psychological Review, 45, 107-133. MacKinnon, M. M. (1972). Adaptation-level theory, anchor theory and the peakshift phenomenon. Journal of Motor Behavior, 4, 1-12. Newlin, R. L., Rodgers, J. P., & Thomas, D. R. (1979). Two determinants of the peak shift in human voluntary stimulus generalization. Perception and Psychophysics, 25, 478-486. Nicholson, J. N., &Gray, J. A. (1971). Behavioral contrast and peak shift in children. British Journal of Psychology, 62, 367-373. Nicholson, J. N., &Gray, J. A. (1972). Peak shift, behavioural contrast and stimulus generalization as related to personality and development in children. British Journal of Psychology, 63, 47-62. Parducci, A. (1974). Contextual effects: A range-frequency analysis. In E. C. Carterette & M. P. Friedman (Eds.), Handbook of perception (Vol. II, pp. 121-141). San Diego, CA: Academic Press. Premack, D. (1978). On the abstractness of human concepts: Why it would be difficult to talk to a pigeon. In S. H. Hulse, H. Fowler, & W. K. Honig (Eds.), Cognitive processes in animal behavior (pp. 423-451). Hillsdale, NJ: Erlbaum. Sams, V. (1990). Context effects in animal psychophysics: A comparative analysis of the chicken's perceptual relativity. European Bulletin of Cognitive Psychology, 10, 475-489. Spence, K. W. (1937). The differential response in animals to stimuli varying within a single dimension. Psychological Review, 44, 430444. Thomas, D. R. (1962). The effects of drive and discrimination training 140 THOMAS, MOOD, MORRISON, AND WIERTELAK on stimulus generalization. Journal of Experimental Psychology, This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. 64, 24-28. Thomas, D. R. (1974). The role of adaptation-level in stimulus generalization. In G. H. Bower (Ed.), The psychology of learning and motivation (Vol. 8, pp. 91-145). San Diego, CA: Academic Press. Thomas, D. R., & Barker, E. (1964). The effects of extinction and central tendency on stimulus generalization in pigeons. Psychonomic Science, 1, 119-120. Thomas, D. R., & Hiss, R. H. (1963). A test of the "Units Hypothesis" employing wavelength generalization in human subjects. Journal of Experimental Psychology, 65, 59-62. Thomas, D. R., & Jones, C. G. (1962). Stimulus generalization as a function of the frame of reference. Journal of Experimental Psychology, 67, 77-80. Thomas, D. R., Strub, H., & Dickson, 1. F., Jr. (1974). Adaptationlevel and the central tendency effect on stimulus generalization. Journal of Experimental Psychology, 103, 466-474. Thomas, D. R., Svinicki, M. D., & Vogt, J. (1973). Adaptation-level as a factor in human discrimination learning and stimulus generalization. Journal of Experimental Psychology, 97, 210-219. Thomas, D. R., & Thomas, D. H. (1974). Stimulus labeling, adaptation-level, and the central tendency shift. Journal of Experimental Psychology, 103, 896-899. White, K. G., & Thomas, D. R. (1979). Postdiscrimination stimulus generalization in humans: An extension of Galizio and Baron. Animal Learning & Behavior, 7, 564-565. Zoeke, B., Sards, V., & Hofer, G. (1988). Psychophysical context effects in chickens (hubbards). The International Journal of Comparative Psychology, I, 167-178. Received May 23, 1990 Revision received August 8, 1990 Accepted August 23, 1990 Low Publication Prices for APA Members and Affiliates Keeping You Up to Date: All APA members (Fellows, Members, and Associates, and Student Affiliates) receive-as part of their annual dues-subscriptions to the American Psychologist and the APA Monitor. High School Teacher and Foreign Affiliates receive subscriptions to the APA Monitor and they can subscribe to the American Psychologist at a significantly reduced rate. In addition, members and affiliates are eligible for savings of up to 60% on other APA journals, as well as significant discounts on subscriptions from cooperating societies and publishers (e.g., the British Psychological Society, the American Sociological Association, and Human Sciences Press). Essential Resources: APA members and affiliates receive special rates for purchases of APA books, including Computer Use in Psychology: A Directory of Software, the Master Lectures, and Journals in Psychology: A Resource Listing for Authors, Other Benefits of Membership: Membership in APA also provides eligibility for low-cost insurance plans covering life; medical and income protection; hospital indemnity; accident and travel; Keogh retirement; office overhead; and student/school, professional, and liability. For more information, write to American Psychological Association, Membership Services, 1200 Seventeenth StreetNW, Washington, DC 20036, USA or call (703) 247-7760 (collect calls cannot be accepted).