Overall levels of variability were generally sensitive to the contingencies in place throughout the experiment and provided some evidence for resurgence of variability of the target dimension, as shown in Figure 2. In the first phase, baseline, points were delivered for varying on the target dimension and repeating on the alternative. As expected, target U-values were significantly higher than alternative U-values, t(46) = 6.89, p < .0001, indicating that participants generally behaved more variably on the target dimension than on the alternative dimension; that is, the rectangles produced were more evenly distributed across response categories for the target dimension than the alternative dimension. In the second phase, alternative, points were delivered for varying on the alternative dimension and repeating on the target dimension. U-values for the target dimension decreased significantly from baseline to alternative, t(46) = 6.95, p < .0001, and conversely, U-values for the alternative dimension increased significantly from baseline to alternative, t(46) = 6.77, p < .0001. Consistent with these results, U-values were significantly higher for the alternative dimension than for the target dimension during the alternative phase, t(46) = 4.777, p < .0001, indicating that the rectangles produced were more evenly distributed across response categories for the alternative dimension than the target dimension. In the third phase, extinction, all points were withheld, and there was evidence for resurgence of reinforced variability of the target dimension. Specifically, U-values for the target dimension increased significantly from alternative to extinction, t(46) = 6.44, p < .0001.

The majority of participants responded in a manner consistent with resurgence of reinforced variability of the target dimension. Figure 3 shows individual subject U-values for the target dimension across phases. Participants’ data were considered to be consistent with resurgence if target U-values decreased from baseline to alternative and increased from alternative to extinction. As shown in Figure 3A, data for 33 of 47 participants (70.2%) demonstrated resurgence according to this definition (Target Pattern A). A second pattern of responding (Target Pattern B), shown in Figure 3B, involved a decrease in target U-values across phases. This pattern was evident for 7 of 47 participants (14.9%) and points to extinction of variability on the target dimension with no resurgence after alternative reinforcement was removed. All other patterns of responding (Target Pattern C) observed in the remaining 7 of 47 participants (14.9%) were considered to be nonsystematic (see Fig. 3C).

There were substantial individual differences in patterns of responding for the alternative dimension. In typical resurgence preparations, alternative responding is not usually a criterion for determining whether resurgence has occurred (e.g., Craig et al., 2017; Sweeney & Shahan, 2013a, b). However, because the target and alternative responses were different dimensions of the same behavior in the current study, it is necessary to analyze responding on the alternative dimension to fully understand the results. Even though target U-values increased from alternative to extinction, indicating resurgence (Fig. 3A), the data for alternative U-values were less clear. Overall, U-values for the alterative dimension did not change significantly from alternative to extinction, t(46) = 1.85, p = 1.0, and U-values for the target and alternative dimensions were not significantly different during extinction, t(46) = 0.30, p = 1.0.

Individual differences in patterns of responding on the alternative dimension are shown in Figure 4. The first pattern of responding (Alternative Pattern A) included data from 14 of 47 participants (29.8%) and is displayed in Figure 4A. For these participants, alternative U-values increased from baseline to alternative and decreased from alternative to extinction. This response pattern supports the finding of resurgence of target variability, because behavior was sensitive to the contingencies in the first two phases and because we only observed an increase in variability for the target dimension in the final phase (i.e., resurgence). A second pattern of responding (Alternative Pattern B) is apparent in the data from 25 of 47 participants (53.2%), shown in Figure 4B. For these participants, alternative U-values increased from baseline to alternative and alternative U-values increased or remained constant from alternative to extinction. Behavior of these participants was sensitive to the contingencies, but U-values for the target and alternative dimensions were both high in the final phase. This outcome is not inconsistent with resurgence, but is at the same time consistent with other explanations for the high alternative U-values (e.g., extinction-induced response variability; see Discussion). Nonsystematic alternative responding (Alternative Pattern C) are shown in Figure 4C; alternative U-values from the remaining 8 of 47 participants (17.0%) were not sensitive to the different contingencies in the first two phases, which precludes resurgence.

In considering individual patterns of responding on both dimensions together, Table 3 shows the number of participants displaying each combination of target and alternative response patterns. Of the 33 participants who demonstrated resurgence of variability on the target dimension (Target Pattern A; see Fig. 3A), 30 of those participants (90.9%) also responded on the alternative dimension in a manner supporting resurgence (i.e., Alternative Patterns A or B; see Figs. 4A and 4B). In other words, data for the majority of participants (30 out of 47; 64.3%) were consistent with resurgence of variability on the target dimension and systematic responding on the alternative dimension.

In the present study, participants earned points for creating rectangles on a computer screen across three experimental phases. In the first phase, baseline, participants earned points for emitting rectangles that varied along a target dimension, size or location (counterbalanced), but were repetitive on the alternative dimension. In the second phase, alternative, participants earned points for emitting rectangles that varied along the alternative dimension but were repetitive on the target dimension. In the third phase, extinction, all reinforcement was suspended. Overall, in baseline, levels of variability for the target dimension were high and levels of variability for the alternative dimension were lower. In the alternative phase, levels of variability decreased for the target dimension and increased for the alternative dimension. The findings from the first two phases replicate the existing literature showing that variability of specific dimensions of a response could be precisely controlled by differential reinforcement contingencies (Ross & Neuringer, 2002). In the final phase of the current experiment, extinction, levels of variability of the target dimension increased for the majority of participants, consistent with resurgence of reinforced variability of the target dimension of behavior. The demonstration of resurgence of reinforced variability of a specific target dimension of behavior in humans is novel, and replicates resurgence of reinforced variability in pigeons (Galizio et al., 2018).

The present findings are consistent with resurgence; that is, the removal of alternative reinforcement during extinction resulted in the reoccurrence of previously extinguished target responding (e.g., Epstein, 1985). Specifically, in baseline, when only variability along the target dimension was reinforced, participants produced rectangles that were relatively variable along the target dimension but relatively repetitive along the alternative dimension. When the contingencies were reversed in the alternative phase, participants’ behavior adapted accordingly (i.e., more variability on the alternative dimension than the target dimension). During extinction, participants tended to return to baseline responding (i.e., increased variability of the target dimension), a resurgence effect. These data provide some evidence for resurgence of reinforced behavioral variability and may support the idea that variability is an operant dimension of behavior.

Even though some of the present findings are consistent with resurgence of reinforced variability of a target dimension of behavior, the results from the final phase of the experiment are mixed. Target U-values increased from the alternative phase to the extinction phase, indicative of resurgence. For some participants, alternative U-values decreased in the extinction phase, a finding clearly in line with selective resurgence of reinforced variability of the target dimension only. However, levels of variability on the alternative dimension remained high for most participants in the final phase. This outcome is not inconsistent with resurgence; in typical resurgence studies, absolute alternative response rates are frequently higher than target response rates (e.g., Craig et al., 2017; Sweeney & Shahan, 2013a, b). However, there are other potential explanations that could account for the high alternative U-values observed for most participants, such as extinction-induced response variability.

This study is a systematic replication and extension of previous research on resurgence of reinforced behavioral variability in pigeons (Galizio et al., 2018). Although there were several procedural differences between the two experiments, such as the subjects (humans versus pigeons), responses (rectangles on a computer screen versus sequences of keypecks in an operant chamber), and variability contingency (relative-frequency threshold versus lag schedule), the results were similar. In both experiments, levels of variability were high in the baseline phase, low in the alternative phase, and high in the extinction phase (i.e., resurgence). However, in both experiments, levels of variability also increased for the alternative response in the extinction phase. Although this result does not preclude a resurgence explanation, it also does not rule out other explanations, such as extinction-induced response variability.

A major drawback of using the traditional resurgence preparation to study relapse of reinforced behavioral variability is the reliance on extinction in the procedure. Behavioral variability may be induced by the removal of reinforcement (i.e., extinction; e.g., Antonitis, 1951). Therefore, the distinction between behavioral variability arising from reinforcement and behavioral variability induced by extinction is important to consider. Extinction can result in high levels of behavioral variability even with no history of reinforcement for specifically behaving variably (e.g., Antonitis, 1951; Eckerman & Lanson, 1969; Jensen et al., 2014; Mechner, 1958; Mechner et al., 1997; Mintz & Notterman, 1965). U-values alone cannot distinguish between reinforced and extinction-induced variability. There are a number of ways that this issue could be addressed in future research. Procedural changes could include the addition of an inactive response (e.g., Craig et al., 2017). If the behavioral variability observed in extinction has been induced by the removal of reinforcement, the subject’s behavior may vary indiscriminately across all responses. If the behavioral variability observed is directed towards the target response and not towards the inactive response, we would have stronger evidence that variability is an operant. There are also potential analytic strategies that could be used in future research to attempt to distinguish between reinforced and extinction-induced variability (e.g., relative frequency distributions; Neuringer et al., 2001). More research is needed to fully address this issue and more effectively differentiate between the contributions of reinforced behavioral variability and extinction-induced response variability.

A possible explanation for the data from participants whose behavior was inconsistent with resurgence on the target (Fig. 3b) and/or alternative dimension (Fig. 4c) is rule-governed behavior. For many participants, U-values changed very little across phases, which indicates that participants may have behaved in accordance with the contingencies or self-imposed rules during baseline but did not adjust their behavior when the contingencies were altered. Research has shown that instructions can significantly impact behavior on variability tasks (e.g., Souza et al., 2012), so we deliberately provided minimal instructions. Our instructions included how to construct a rectangle, that the goal was to earn as many points as possible, and, between tasks, that they would be playing a different version of the same game. Nevertheless, it is possible that the wording of these instructions may have prompted participants to create their own rules, which could have impacted their behavior. This finding is corroborated by self-report measures collected in the demographic questionnaire. As shown in Table 1, most participants reported that they thought the purpose of the variability task was to make a particular “correct” rectangle at any given time, and many reported that they were responding according to particular patterns. Only one participant reported that the task was about responding variably. Even though accurate description of the programmed contingencies is not necessary to satisfy those contingencies (Hefferline et al., 1959), these results indicate that at least some of the participants may have ultimately been primarily responding to self-imposed rules that incidentally satisfied the experimental contingencies, rather than responding to the contingencies themselves (Baron & Galizio, 1983; Galizio, 1979).

One limitation of the present study is that our primary dependent measure was U-value, which has shortcomings when applied to the study of reinforced behavioral variability. U-value measures variability on a global level and cannot account for the specific responses emitted. The utility of U-value as a measure of variability has recently been questioned, but adequate alternatives have not yet been well established (Kong et al., 2017). Future research should be aimed at investigating more molecular methods of analyzing behavioral variability, and especially methods to quantify how variability changes across conditions.

Given that we used a resurgence preparation to examine relapse of reinforced behavioral variability, there are other issues to consider. For example, a key difference between our preparation and the typical resurgence paradigm is the available response options throughout each phase (for an over-view, see Shahan & Sweeney, 2011). For example, in a typical resurgence experiment with rats, a single response option, the target, is made available during baseline (e.g., lever press). In the alternative phase, the alternative response option is made available for the first time (e.g., chain pull). Importantly, the target and alternative responses are mutually exclusive. That is, the rat cannot press the lever and pull the chain at the same time.

Conversely, in the present study, a single rectangle could be categorized by its size and its location, meaning that the target and alternative responses were available simultaneously throughout the study, and thus never mutually exclusive. We attempted to control for this important procedural difference by altering the contingencies to make the target and alternative responses more distinct. As stated in the Method, a rectangle resulted in a point only if it satisfied a threshold contingency for the dimension currently producing points and a reverse threshold contingency on the other dimension. For example, in baseline, points were only delivered for rectangles that were sufficiently variable on the target dimension (e.g., size) and also were sufficiently repetitive on the alternative dimension (e.g., location). Points were never delivered for high levels of variability on both dimensions simultaneously. That said, a limitation of the present study is that the two response dimensions were not truly mutually exclusive, as they are in most resurgence studies, and could co-occur during extinction.

Despite these limitations, the finding in the present study of evidence for resurgence of reinforced behavioral variability in humans has important theoretical and practical implications. At a theoretical level, demonstrating relapse of reinforced behavioral variability provides further evidence that variability is an operant dimension of behavior. Resurgence of reinforced behavioral variability has been demonstrated in pigeons (Galizio et al., 2018), but the present experiment will be the first published study¹ to directly examine and demonstrate resurgence of variability in humans. However, the results also elucidate the difficulty in studying reinforced behavioral variability in a relapse preparation, especially due to the difficulties of parsing extinction-induced variability from relapse of reinforced behavioral variability.

On a practical level, these findings may also inform clinical interventions in applied settings. For example, individuals diagnosed with ASD experience a number of behavioral deficits, including the tendency to behave repetitively (APA, 2013). Even when it would be more beneficial to vary their responses, individuals with ASD often engage in stereotypy. For example, when playing with blocks, a peer may make many variable structures, but the child with ASD may construct the same arrangement of blocks repeatedly. Such behavior may limit the degree to which the two children will engage in social interaction. Variability training has been shown to be beneficial to individuals with ASD in facilitating social interactions and allowing individuals to more effectively contact reinforcement in various settings (e.g., Brodhead et al., 2016; Contreras & Betz, 2016; Goetz & Baer, 1973; Lee & Sturmey, 2006, 2014). Unfortunately, such training is likely to be subject to lapses in treatment fidelity, which makes the investigation of resurgence useful. The finding that reinforced behavioral variability is susceptible to relapse may inform both theoretical interpretations and treatment strategies in clinical settings.