Translational evaluation of on/off alternative reinforcement cycling

Abstract

Cycling between the availability and unavailability of reinforcement for alternative responding has successfully reduced resurgence in basic laboratory evaluations, but this approach represents a marked departure from current standards of care when treating problem behavior, warranting careful translation before its use clinically. Therefore, with extinction arranged for target responding across groups in Phase 2, we evaluated the effects of cycling between the availability and unavailability of differential reinforcement of alternative behavior (DRA) using a computer-based task with adult humans recruited through Amazon MTurk. Two control groups experienced constant DRA in Phase 2, with one group experiencing a dense DRA schedule and another group experiencing a lean DRA schedule. The cycling DRA group tended to show greater reductions in target responding and improved discrimination in Phase 2 and less target responding across Phases 2 and 3 than the lean DRA and dense DRA groups. These preliminary findings suggest that on/off DRA cycling procedures may produce more desirable treatment outcomes than constant DRA without producing negative side effects; however, further research is needed to confirm these possibilities.

Resurgence is a form of relapse in which a previously suppressed response recurs due to a worsening of current reinforcement conditions (Lattal et al., 2017). Resurgence is highly prevalent in clinical settings (Briggs et al., 2018; Falligant, Chin, et al., 2022; Haney et al., 2022; Kranak & Falligant, 2021; Mitteer et al., 2022; Muething et al., 2021)—a finding that has spurred new basic, translational, and applied research evaluating the variables influencing its occurrence as well as methods for its mitigation (see Kimball et al., 2023, for a recent review). Translating findings on resurgence from basic laboratories to applied settings has shown promise (Fisher, Greer, Craig, et al., 2018) and has informed the development of effective resurgence-mitigation strategies in the clinic (e.g., Fisher et al., 2020; Fisher, Greer, Fuhrman, et al., 2018; Fuhrman et al., 2016).

On/off alternative reinforcement cycling described most recently by Shahan et al. (2020) is a promising resurgence-mitigation strategy that warrants similar translation. Shahan et al. evaluated the effects of rapidly alternating periods of availability and unavailability of differential reinforcement of alternative behavior (DRA) on resurgence with rats. During treatment (Phase 2), groups of rats experienced DRA that was “constant on” or “on/off.” Across groups, target responding remained on extinction in Phase 2. For the constant on groups, alternative responding consistently produced reinforcement in Phase 2, and constant on groups experienced different durations of treatment (i.e., 3–31 days). In the on/off group, alternative responding produced reinforcement during odd-numbered sessions and extinction on even-numbered sessions in Phase 2. Resurgence for the on/off group was significantly lower than that for the constant-on groups at each treatment duration compared. Similar results have been obtained in two earlier investigations of on/off DRA cycling (Schepers & Bouton, 2015; Trask et al., 2018).

Despite these impressive results, these procedures differ from recommended clinical practices in several important ways, which may dissuade researchers from attempting to translate this research to more applied settings, populations, and behaviors. First, on/off DRA cycling involves early and frequent exposure to extinction when initiating treatment, which departs from current standards of care that often limit extinction exposure early on when treating problem behavior (see Fisher, Greer, Mitteer, et al., 2018, and Fisher et al., 2023, for data relevant to these two approaches). Second, extinction is rarely used as the sole intervention component when treating problem behavior (see Norris et al., in press, for discussion), meaning that entire sessions without the availability of alternative reinforcement have become uncommon in applied behavior analysis research and practice. Third, discrimination training in the clinic is often accomplished using experimenter-selected, condition-correlated stimuli (e.g., colored wristbands to signal the differential availability of alternative reinforcement), whereas on/off DRA cycling employs no such stimuli and likely trains a different type of discrimination about the underlying contingencies in effect during treatment (see Shahan et al., 2020, for elaboration). For these reasons, on/off DRA cycling represents a marked departure from current standards of care for treating problem behavior, yet these same departures, paired with the available basic research supporting the efficacy of this approach for mitigating resurgence, simultaneously make its potential translation to the clinic all the more intriguing.

For applied researchers, the aforementioned departures from standards of care may be troublesome because these recommended clinical practices are designed, in part, to decrease the potentially deleterious effects of extinction procedures. For example, Lerman et al. (1999) conducted a record review of 41 participants whose self-injurious behavior had been reduced using either (a) extinction or (b) extinction in conjunction with another treatment procedure (e.g., differential reinforcement, noncontingent reinforcement). They found that 62% of the cases demonstrated an initial increase in SIB or aggression (i.e., an extinction burst) when extinction was implemented in isolation, whereas extinction bursts occurred for only 20% of cases when extinction was implemented in conjunction with other treatment procedures. This has led to the recommendation that clinicians should not implement extinction without other treatment components because it increases the probability of bursts of destructive behaviors (e.g., SIB, aggression), which could lead to physical injuries. Thus, applied researchers may be wary about evaluating the effects of procedures like on/off DRA cycling, which represents a significant departure from treatments designed to mitigate the risk of extinction bursts.

When evaluating the risk of physical injury that may be associated with on/off DRA cycling, clinicians will also be interested in evaluating the total number of target responses a person emits throughout the implementation of the procedure. For applied researchers, each target response may be analogous to a response that has the potential for causing physical injury (e.g., aggression and self-injury). Thus, applied researchers are interested in identifying procedures that lead to the fewest number of target responses overall because this mitigates the overall risk of harm associated with conducting the procedure. Thus, it is important to evaluate not only the effects of on/off DRA cycling produces on extinction bursting but also its effects on cumulative target responding.

Similarly, applied researchers may have concerns about the effects of on/off DRA cycling on the acquisition and maintenance of alternative responses. For example, clinicians who assess and treat severe destructive behavior may be concerned that putting a client’s newly acquired response on extinction would impede response acquisition, fail to maintain the response over time, or result in undiscriminated responding. In fact, clinicians often detect such clinical issues by evaluating patients’ discrimination indices (i.e., the proportion of alternative responses emitted when reinforcement is available relative to their overall target and alternative responding) and making clinical decisions based on these data (e.g., Greer et al., 2016). Thus, it is also necessary to evaluate how on/off DRA cycling affects responding in ways that are more relevant to clinical practice, such as evaluating its effects on participants’ discrimination indices.

Conducting experiments with computer-based tasks is a useful approach for extending research to human participants while minimizing potential risks. There is a growing body of research demonstrating that basic behavioral phenomena can be translated using computer-based tasks completed by adults recruited through crowdsourcing websites. For example, Robinson and Kelley (2020) demonstrated that renewal and resurgence phenomena can be replicated using a computer-based task with participants recruited through Amazon’s MTurk platform. Other studies (e.g., Podlesnik et al., 2022; Ritchey et al., 2023; Smith & Greer, 2022a) have demonstrated that human behavior on these tasks can be described by quantitative models like Resurgence as choice in context (RaC²; Shahan et al., 2020), which has been used to accurately predict the behavior of both nonhuman animals (e.g., Shahan et al., 2020) and clinical populations (e.g., Falligant, Hagopian, et al., 2022; Shahan & Greer, 2021). Thus, human behavior on computer-based tasks may reflect the same basic behavioral processes that govern the behavior of animals in the laboratory and that of clinical populations.

The purpose of this research was to extend Shahan et al. (2020) by evaluating whether on/off DRA cycling mitigates resurgence of target responding in human participants using a computer-based task that minimizes the potential risks of harm that may be associated with repeated exposures to extinction in the clinic. To aid this translation, the present study also employed a number of procedural modifications directed at more closely mimicking the procedures used during the treatment of problem behavior. These included shorter session durations (i.e., 2 min including reinforcement time in this experiment instead of 30 min excluding reinforcement time in Shahan et al., 2020), fewer cycles of on/off DRA (i.e., 6 cycles instead of 16 cycles), and less time required for the evaluation (i.e., about 45 min instead of 2,130 min excluding pretraining and reinforcement time). Finally, the present study also evaluated participants’ cumulative target responses and discrimination indices because of the importance of these metrics for clinical practice.

METHOD

Participants and setting

We recruited participants through Amazon’s MTurk platform using the qualifications feature to recruit participants who (a) had greater than 99% task approval ratings, (b) had completed at least 50 tasks, (c) lived in the United States, and (d) had not previously completed an experiment conducted by our lab. Notably, workers on MTurk must be at least 18 years old to register, so all our participants were at least 18 years old. Participants used a personal electronic device to complete the task in any location where they were able to access the MTurk platform. We posted the human intelligence task (HIT) with a reward of $0.10 for completion, and we informed participants they could receive up to $5 per hour based on their performance during the task, which we paid using MTurk’s “bonus” compensation feature.

We conducted a power analysis based on the results of a pilot study to determine how many participants we should recruit to detect between-group differences in resurgence. In the pilot study, we included two experimental groups similar to the cycling DRA and dense DRA groups in this experiment. There was a large difference in target responding at the beginning of Phase 3 between these two groups (Cohen’s d = .98). Based on this effect size, our power analysis indicated we would need nine participants in each of our three experimental groups to obtain .80 power at α = .05 for detecting differences in resurgence at the beginning of Phase 3. We conservatively increased our recruitment to 16 participants per group for two reasons. First, our pilot study did not include a lean DRA group, which is predicted to show less resurgence than dense DRA (e.g., Craig et al., 2016; Craig & Shahan, 2016; Sweeney & Shahan, 2013), so we increased our N to provide additional power to detect this smaller predicted difference. Second, we did not want to have fewer participants than the smallest n used in previous behavior-analytic research on relapse phenomena using participants recruited through MTurk. We identified n = 16 as the smallest value previously published (i.e., in Smith & Greer, 2022a). Collectively, we determined we would need 16 participants per group.

Variables and design

We evaluated the effects of three different DRA procedures during Phase 2 of a three-phase resurgence evaluation using a between-groups design. One group experienced dense alternative reinforcement (dense DRA), a second group experienced lean alternative reinforcement (lean DRA), and a third group experienced on/off alternative reinforcement cycling (cycling DRA). Participants were randomly selected into one of the groups upon enrollment in the experiment.

The primary dependent variables were correct completion of target and alternative response chains. Both chains involved (a) clicking a shape with the computer cursor, (b) clicking a free-text field and typing a number that solved the arithmetic problem displayed on the screen, and (c) clicking the submit button. Target and alternative response chains were differentiated only by the shape of the button in the initial link (i.e., star for target, square for alternative) and the side of the interface where the buttons appeared semirandomly (i.e., left half of the interface for the target, right half of the interface for the alternative).

We also collected frequency data on incorrect target and incorrect alternative responses, each of which were the same as their respective correct response chains, except the number typed during the second link of the chain did not solve the arithmetic problem.

General procedures

Our procedures were similar to those described by Smith and Greer (2022a), and we validated the functionality of the software using methods similar to those described by Smith and Greer (2022b). After participants accessed the experiment through the MTurk platform, the experimental interface presented the informed consent narrative to participants, followed by a consent quiz. Participants had to answer all questions correctly to pass the quiz. If participants failed, they returned to the consent narrative to review the narrative. If participants failed the quiz three times, they were not permitted to complete the experiment. If participants passed the consent quiz within their first three attempts, participants completed a task to adjust the zoom on their Internet browser to ensure they would be able see the entire experimental interface without scrolling. The experimental software then presented the following instructions:

Your task for the experiment will require you to use your cursor and keyboard to input the correct answers to math problems that will be presented on your computer screen. On your screen, one or two different shapes will appear and move around the screen randomly. First, you must click one of the shapes. Next, a math problem will appear, and you will need to type the correct number that answers the math question and click the “SUBMIT” button. You might collect money when you submit the correct answer. The goal of the game is to get as much money as possible. Your compensation will be directly tied to how much money you collect. If you collect more money during the experiment, you will be paid more money at the end of the experiment. The maximum amount that you can earn is $5. Do NOT press the back button on your browser or refresh your browser window during the experiment, it will start the experiment over again from the beginning.

After the instructions, participants completed a second comprehension quiz that was similar to the consent quiz, except there was no limit on the number of times the participant could complete the comprehension quiz. Participants proceeded to the experiment after answering all the quiz questions correctly.

Resurgence evaluation

All participants experienced the same resurgence-evaluation conditions, except for the DRA procedures in Phase 2. Participants experienced 2-min sessions throughout the resurgence evaluation. All sessions were separated by a 5-s blackout. Across all sessions and phases, the experimental software displayed the total amount of money each participant had earned throughout the experiment.

Phase 1

The target response button in the shape of a star was present and moved randomly around the left half of the display after each response. When the participant clicked the target response button, an addition problem (e.g., 3 + 9) appeared along with a field to type the answer and a button to click to submit the answer. Clicking the button to submit the answer made the math problem disappear. If the participant typed the correct answer into the answer field prior to clicking the submit button and a reinforcer was set up, clicking the button produced reinforcement in the form of increasing the participant’s money total by $.015. To increase the discriminability of each monetary deposit, stars corresponding to the initial link of the target response chain flashed around the money counter. The first correct target response in each session produced reinforcement, and money was delivered according to a variable-interval (VI) 15-s schedule for the remainder of the session. The software produced the VI 15-s schedule by randomly selecting a number between 10 and 20 s to use as the next schedule interval after each reinforcer delivery. Incorrect responses were placed on extinction (i.e., no money was deposited, and no star flashed around the money counter). The alternative response was unavailable in Phase 1. Phase 1 lasted five sessions (i.e., 10 min).

Phase 2

Correct and incorrect target responses were placed on extinction. In addition to the target response button, an alternative response button (i.e., a square) appeared and moved around the right side of the display following each response. Clicking the square made an addition problem appear, and submitting a correct response when a reinforcer was set up produced a $.015 increase in the participant’s money counter, and squares flashed around the total to signal the monetary deposit. Incorrect alternative responses were placed on extinction. Phase 2 lasted 11 sessions (i.e., 22 min). Reinforcement for correct alternative responses varied across experimental groups, described below.

Dense DRA

The first correct alternative response in each session produced reinforcement, and money was deposited for correct alternative responses according to a VI 5-s schedule thereafter. The software produced the VI 5-s schedule by randomly selecting a number between 2.5 and 7.5 s to use as the next schedule interval after each reinforcer delivery. Alternative reinforcement was available in every session of Phase 2.

Cycling DRA

Alternative reinforcement was available only during even-numbered sessions of Phase 2. When reinforcement was available, the first correct alternative response in each session produced reinforcement and money was deposited according to a VI 5-s schedule thereafter. The software produced the VI 5-s schedule by randomly selecting a number between 2.5 and 7.5 s to use as the next schedule interval after each reinforcer delivery. Alternative reinforcement was unavailable during odd-numbered sessions of Phase 2.

Lean DRA

The first correct alternative response in each session produced reinforcement, and money was deposited according to a VI 15-s (interval range, 10–20 s) schedule thereafter. The software produced the VI 15-s schedule by randomly selecting a number between 10–20 s to use as the next schedule interval after each reinforcer delivery. Alternative reinforcement was available in every session of Phase 2.

We included a lean DRA group to provide a comparison group for the cycling DRA group that included fewer reinforcer deliveries throughout Phase 2. Previous research has shown that the overall rate of reinforcement in Phase 2 affects resurgence, with denser reinforcement rates tending to produce more resurgence than leaner ones (e.g., Craig et al., 2016; Craig & Shahan, 2016; Sweeney & Shahan, 2013). Although Shahan et al. (2020) observed less resurgence in their on/off group than in their constant-on groups, those differences could have been due to their on/off group obtaining fewer alternative reinforcers in Phase 2. Therefore, we included a lean DRA group to help rule out this possibility and better isolate the effects of cycling between alternative reinforcer availability and unavailability.

We selected a VI 15-s schedule for the lean DRA group to approximate the predicted reinforcement rates for the cycling DRA group based on data from Smith and Greer (2022a) and data we collected during a pilot study. During this experimental task, participants rarely obtain the maximum reinforcement available on a given reinforcement schedule because (a) participants may not fully discriminate the contingencies in place and (b) participants are limited by how quickly they can complete each response chain. Thus, we used data from Smith and Greer (2022a) and pilot data to calculate proportions of reinforcers participants typically obtain as a function of the programmed VI schedule interval. Analysis of these data suggested a VI 15-s schedule for the lean DRA group would approximate (and not exceed) the obtained reinforcement in the cycling DRA group. This ensured that the lean DRA group could serve as a control group for the effects of different obtained reinforcement rates during Phase 2.

Phase 3

Target and alternative response buttons remained available, and clicking them continued to produce addition problems; however, completing either response chain did not produce reinforcement. Phase 3 lasted six sessions (i.e., 12 min).

Debriefing

Although participants were told their compensation was related to their performance during the experiment, we paid all participants the maximum possible rate of $5 to ensure equity across participants. We explained this procedure to participants with a debriefing script at the end of the experiment and offered participants the opportunity to withdraw consent at this time. No participant withdrew consent.

Comparison with the procedures of Smith and Greer (2022a)

Our procedures were identical to those of Smith and Greer (2022a), with the following exceptions. First, the initial links in our target and alternative response chains were differentiated by shapes (i.e., star, square) rather than colors (i.e., blue, yellow). We made this change to increase the accessibility of the task for individuals with color-blindness. Second, we divided the task into discrete sessions by adding 5-s blackout periods every 2 min. We made this change to create discriminable sessions within the experimental task to (a) resemble cycling procedures in basic laboratories where contingencies change across (not within) sessions and (b) increase the discriminability of contingency changes. Third, we delivered reinforcers in US dollars instead of points. We made this change as an attempt to increase the salience of the reinforcement contingencies throughout the experiment.

Data analysis

Our data analysis included several critical comparisons that are relevant for translating the effects of cycling DRA to applied settings. First, we evaluated whether (a) resurgence occurred, (b) target response rates differed across groups at the end of Phase 2, and (c) target response rates differed across groups at the beginning of Phase 3. Next, we evaluated whether the cumulative number of target responses differed across experimental groups. Finally, we evaluated whether discrimination indices differed across experimental groups at the end of Phase 2.

RESULTS

We recruited participants until we obtained 16 viable data sets for each experimental group (i.e., 48 participants total). We obtained 57 total data sets and excluded nine (i.e., 84% retention). We deemed data sets unviable if the participant engaged in zero responses for two or more consecutive sessions (i.e., ≥4 min) in Phases 1 or 2, which demonstrated disengagement from the experiment. This occurred for one participant in the dense DRA group and for three participants in the lean DRA group. We also excluded data sets if the participant (a) engaged in zero target responses across Phases 2 and 3 and (b) showed no decrease in alternative responding during Phase 3 because this demonstrated insensitivity to the contingency change from Phase 2 to Phase 3. This occurred for three participants, one participant in the dense DRA, cycling DRA, and lean DRA groups, respectively. A total of 56 prospective participants failed to pass the consent quiz after three attempts. These individuals were excluded from participating. We also collected and analyzed data on traffic to the website hosting our experiment, which did not show any anomalies.

Table 1 displays the mean target and alternative reinforcer deliveries in the last two sessions of Phase 1, the last two sessions of Phase 2, and all of Phase 3. We included data from two sessions to ensure that the data reflect one of each type of session experienced by the cycling DRA group in Phase 2 (i.e., one session with alternative reinforcement available, one session with alternative reinforcement unavailable). Obtained target reinforcement was similar across groups at the end of Phase 1. At the end of Phase 2, obtained alternative reinforcement was highest in the dense DRA group, followed by the cycling DRA group, and lowest in the lean DRA group. This shows that our schedule manipulations successfully produced different reinforcement rates across groups to help rule out obtained reinforcement rate as the primary variable contributing to the effects of cycling DRA procedures. It is worth noting that the standard deviation for obtained reinforcement in Phase 2 for the cycling DRA group is larger than that of the other groups; however, obtained reinforcement was 0 for all participants in this group during one of these sessions (i.e., when alternative reinforcement cycled “off”), which contributed to the larger spread in these data.

TABLE 1.

Obtained reinforcement

	Phase 1		Phase 2
Group	Target	Alternative	Target	Alternative
Cycling DRA	7.16 (.68)	0 (0)	0 (0)	7.53 (8.00)
Dense DRA	6.69 (.74)	0 (0)	0 (0)	11.72 (4.99)
Lean DRA	7.03 (.69)	0 (0)	0 (0)	5.28 (2.07)

Note. Data represent the mean target and alternative reinforcer deliveries in the last two sessions of Phase 1 and the last two sessions of Phase 2. Data in parentheses represent standard deviations.

The top panel of Figure 1 displays the group means of target responding in each minute of the experiment (please see Supplemental Information for data from individual participants). We observed similar patterns of target responding across groups in Phase 1. We conducted a one-way analysis of variance (ANOVA) on the data from the last session of Phase 1 to evaluate whether differences in target responding at other points during the experiment may be due to differences in baseline response rates across groups. This ANOVA did not show significant differences across groups, F(2, 45) = .96, p = .39.

FIGURE 1.

Mean responses per session. The top and bottom panels represent mean target and alternative responses during each session of the experiment, respectively. Error bars represent standard error of the mean.

In Phase 2, target responding decreased across all experimental groups. Target responding decreased more rapidly at the beginning of Phase 2 in the dense DRA and cycling DRA groups relative to the lean DRA group. Notably, target responding in the cycling DRA group fluctuated throughout Phase 2 in a way that was consistent with the availability and unavailability of alternative reinforcement: Target responses were lower in even-numbered sessions when alternative reinforcement was available and higher in odd-numbered sessions when alternative reinforcement was unavailable. Despite these increases in target responding during odd-numbered sessions, there were only two sessions when target responding in the cycling DRA exceeded target responding in another group (i.e., target responding was greater in the cycling DRA group relative to the dense DRA group in sessions 7 and 13). Although both occurrences were when alternative reinforcement cycled “off,” target responding in the cycling DRA group remained in a range similar to that in the dense DRA group and lower than that in the lean DRA group in both instances. By the end of Phase 2, target responding occurred at a lower rate in the cycling DRA group relative to the dense DRA and lean DRA groups. Target responding in the dense DRA and lean DRA groups occurred at similar rates at the end of Phase 2.

The bottom panel of Figure 1 displays group means of alternative responding in each session of the experiment. Alternative responding increased during Phase 2 across all groups, and alternative responding was similar across the dense and lean DRA groups at the end of Phase 2. Alternative responding fluctuated throughout Phase 2 in a way that was consistent with the availability and unavailability of alternative reinforcement: Alternative responses were higher in even-numbered sessions when alternative reinforcement was available and lower in odd-numbered sessions when alternative reinforcement was unavailable. Alternative responding was greater in the cycling DRA group than in the dense DRA and lean DRA groups at the end of Phase 2.

In Phase 3, target responding increased in the cycling DRA and lean DRA groups, but target responding increased to a lesser extent in the dense DRA group. Alternative responding decreased across all experimental groups in Phase 3. Notably, neither target nor alternative responding approached 0 for any group during Phase 3.

Figure 2 displays individual data from the last two sessions of Phase 2 (left two panels) and the first two sessions of Phase 3 (right two panels), which is the critical point in the experiment to determine the effects of the different contingencies throughout Phase 2. There was a tendency for target responding to be lower and less variable in the cycling DRA group than in the other groups at the end of Phase 2 and the beginning of Phase 3. Notably, target responding tended to be lower in the cycling DRA group at the end of Phase 2 even though alternative responding was on extinction in the second to last session of Phase 2 (left-most panel of Figure 2).

FIGURE 2.

Target response frequency at the end of Phase 2 and the beginning of Phase 3. Each panel displays individual participants’ target response frequencies in a single session of the experiment. Sessions 15 and 16 were the last two sessions in Phase 2. Sessions 17 and 18 were the first two session in Phase 3. Horizontal lines represent group means and error bars represent standard error of the mean. Please note the break in the y-axis in each panel.

We supplemented these visual analyses with statistical analyses. We could not conduct a mixed-model ANOVA to evaluate resurgence, group differences, and interactions during the transition from the end of Phase 2 to the beginning of Phase 3 because our data violated the assumption of normality of residuals. To avoid violating the assumptions of this test, we conducted three separate statistical analyses to evaluate whether (a) resurgence occurred, (b) target responding differed between groups at the end of Phase 2, and (c) target responding differed between groups at the beginning of Phase 3. To account for Type 1 error rate, we conducted a Bonferroni alpha correction: we set α = .05 and divided this by the number of statistical tests (i.e., three) used to evaluate these data such that α = .05/3 = .017.

To evaluate whether resurgence occurred, we compared target responding in the last session of Phase 2 with responding in the first session of Phase 3 for all participants across all groups. These data violated the assumption of normality of residuals, so we needed to conduct a nonparametric test. To evaluate whether our data met the assumptions for the Wilcoxon matched-pairs signed-rank test, we conducted a D’Agostino–Pearson test to evaluate whether the differences in the pairs (i.e., each participant’s data from the last session of Phase 2 to the first session of Phase 3) were distributed symmetrically around their median. The data passed this normality test (K² = 1.501, p = .4722), suggesting that the Wilcoxon matched-pairs signed-rank test was an appropriate statistical test. The Wilcoxon matched-pairs signed-rank test indicated that target responding in the first session of Phase 3 (Mdn = 3.50) was significantly greater than in the last session of Phase 2 (Mdn = 1.50), W = 259.0, p = .0064, r = .67. This generally confirms that resurgence occurred on an experiment-wide basis.

Next, we evaluated whether target responding differed between groups in the last session of Phase 2. These data also violated the assumption of normality of residuals, so we conducted a nonparametric Kruskal–Wallis test. Although this test showed a moderate effect of group assignment on target responding in the last session of Phase 2 (η² = .1032), this test did not show significant differences when the Bonferroni alpha correction (H = 6.642, p = .0361) was applied. Although this statistical test did not suggest a significant difference between groups in the last session of Phase 2, there appears to be an important, nominal tendency for target responding to be lower in the cycling DRA group when the last two sessions of Phase 2 are evaluated visually (as depicted in the left two panels of Figure 2).

We also evaluated whether target responding differed between groups in the first session of Phase 3. These data also violated the assumption of normality of residuals, so we conducted a nonparametric Kruskal–Wallis test, which showed a small effect size (η² = .0364) in the first session of Phase 3, and this effect was not statistically significant (H = 3.636, p = .1624).

The top panel of Figure 3 displays mean cumulative target responses across Phases 2 and 3 for all experimental groups. Although Figure 3 does not further elucidate whether resurgence occurred differentially across groups, the critical purpose of Figure 3 is to display that cumulative target responses in the cycling DRA group never exceeded the cumulative target responses in the dense DRA or lean DRA groups. Further, the cumulative target responses for the dense and lean DRA groups increased at a faster rate than the cycling DRA group throughout Phase 2. This resulted in large between-group differences in the total number of target responses that occurred throughout Phases 2 and 3, which may have important clinical implications for target behavior reduction per se.

FIGURE 3.

Mean cumulative target responses and discrimination indices. The top panel represents mean cumulative target responses during Phases 2 and 3 of the experiment. The bottom panel represents mean discrimination indices during Phase 2 of the experiment. Error bars represent standard error of the mean.

We were unable to conduct a mixed-model (Session × Group) ANOVA on cumulative target responding due to heteroscedasticity of variances (the variance of cumulative target responses increased throughout the experiment). Further, data transformations (e.g., logarithmic) did not resolve the issue of heteroscedasticity of variances. In lieu of significance testing, we calculated Cohen’s d to evaluate the size of the effect of different experimental groups on cumulative target responding at the end of Phases 2 and 3. At the end of Phase 2 (i.e., Session 16), there was a medium effect between the cycling DRA and dense DRA groups on cumulative target responding (Cohen’s d = .4967), a large effect between the cycling DRA and lean DRA groups (d = .8304), and a small effect between the dense DRA and lean DRA groups (d = .2507). At the end of Phase 3, there was a small effect between the cycling DRA and dense DRA (d = .3623), a large effect between the cycling DRA and lean DRA (d = .7625), and a small effect between dense DRA and lean DRA (d = .3242).

The bottom panel of Figure 3 displays mean discrimination indices during Phase 2 of the experiment. Discrimination indices are an important metric used by applied researchers because they quantify the extent to which participants learn to engage in “correct” alternative responses (i.e., when reinforcement is available). We calculated discrimination indices for each participant by dividing the number of alternative responses in each session by the sum of target and alternative responses in that session. Next, we averaged these indices within each group to identify the group mean for each session. Discrimination indices increased across Phase 2 for all experimental groups. Discrimination indices were similar in the dense DRA and lean DRA groups throughout Phase 2. In the cycling DRA group, discrimination indices fluctuated with the availability and unavailability of alternative reinforcement. Despite these fluctuations, there were only two sessions when the discrimination indices of the cycling DRA group were lower than the discrimination indices of the other groups (i.e., Sessions 7 and 13). In general, discrimination indices increased more rapidly in the cycling DRA group and were greater than the dense DRA and lean DRA groups by the end of Phase 2.

We supplemented this visual analysis with a statistical analysis evaluating whether discrimination indices differed across groups by the end of Phase 2. The discrimination indices in the last session of Phase 2 violated the assumption of normality of residuals, so we conducted a nonparametric Kruskal–Wallis test, which showed significant moderate differences across experimental groups (H = 6.828, p = .0329, η² = .1073). This generally confirms that the different contingencies across groups in Phase 2 affected participant’s discrimination indices.

DISCUSSION

We compared the effects of on/off DRA cycling with conditions in which alternative reinforcement remained available during treatment with human participants completing a computer-based task. We found that, compared with both dense DRA and lean DRA, cycling DRA produced several beneficial effects that could have important implications for applied researchers. Specifically, the cycling DRA group tended to show greater reductions in target responding and improved discrimination in Phase 2 and less target responding across Phases 2 and 3 than did the lean DRA and dense DRA groups.

Most important of these findings is that target responding throughout Phase 2 remained as low or lower in the cycling DRA group than in the dense DRA and lean DRA groups. Although cycling DRA produced fluctuating rates of target responding throughout Phase 2 (see top panel of Figure 1), target responding in the cycling DRA group was only greater than the dense DRA group in two sessions (i.e., Sessions 7 and 13) and was never greater than the lean DRA group. Further, target responding appears to have been lower in the last two sessions of Phase 2 in the cycling DRA group than in the dense DRA and lean DRA groups (see left two panels of Figure 2). Moreover, cumulative target responding in Phase 2 for the cycling DRA group remained lower than that of the other groups despite the repeated removal of alternative reinforcement throughout Phase 2 (see top panel of Figure 3). Differences in cumulative target responding across groups actually increased throughout Phase 2, and these differences remained evident throughout Phase 3.

Identifying procedures that lead to fewer cumulative target responses could have considerable applied value regardless of the extent to which resurgence occurs. For example, in a clinic setting, if the target response is head banging against hard surfaces, it is important to identify procedures that will lead to fewer cumulative responses compared with other procedures. Collectively, our findings suggest that on/off DRA cycling is a promising procedure for potentially producing greater decreases in target responding when compared with at least some conditions in which DRA remains “on.” If replicated in future research, the suppressive effects of on/off DRA cycling on target responding could be clinically important.

Shahan et al. (2020) suggested that on/off DRA cycling may lead to more complete suppression of target responding because organisms learn to discriminate the availability and unavailability of reinforcers more rapidly. In short, repeated exposures to extinction for alternative responding will lead organisms to learn that (a) reinforcer deliveries signal the availability of additional reinforcers and (b) the absence of reinforcer deliveries signals subsequent reinforcer unavailability. Our analysis of discrimination indices provides initial evidence that on/off DRA cycling may produce better discrimination between target and alternative responding in treatment. The bottom panel of Figure 3 shows that discrimination indices tended to be higher in the cycling DRA group throughout the majority of Phase 2, suggesting that on/off DRA cycling procedures may enhance discrimination of the contingencies in effect based on presence and absence of reinforcer deliveries.

An additional benefit of on/off DRA cycling during treatment may be to mitigate resurgence. In this experiment, nominally less target responding occurred in the cycling DRA group at the beginning of Phase 3 than in the dense DRA and lean DRA groups. Although cycling DRA did not produce statistically significant lower target responding at the beginning of Phase 3 based on a statistical analysis of a single session, visual analysis of two sessions at the beginning of Phase 3 (i.e., the right two panels of Figure 2) suggests that cycling DRA produced less target responding across 4 min of extinction. For example, in the cycling DRA group, there was a greater number of sessions without any target responses (i.e., 14 data points at 0) than in the dense DRA or lean DRA groups (i.e., 8 and 5, respectively). Similarly, there was much less variability in target responding in the first two sessions of Phase 3 in the cycling DRA group than in the dense DRA and lean DRA groups. Thus, although the difference in target response rates was not statistically significant between cycling DRA and the other experimental groups in the first session of Phase 3, there is other evidence suggesting that cycling DRA may produce differences in responding that could be clinically meaningful and should be evaluated further.

The results of this experiment should be considered in relation to its limitations. One limitation is that our experiment did not replicate the effects of alternative-reinforcement rate on resurgence. Previous research has shown that dense alternative reinforcement rates tend to produce more resurgence when reinforcement is suspended (e.g., Craig et al., 2016; Craig & Shahan, 2016; Sweeney & Shahan, 2013). The results of these previous experiments suggest that our dense DRA group would demonstrate greater resurgence than the lean DRA group; however, our lean DRA group demonstrated greater resurgence than the dense DRA group. This observation may have been due to the relatively dense schedules of reinforcement programmed across experimental groups in the present study. For example, by the end of Phase 2 in this experiment, 53.74% and 25.37% of alternative responses produced reinforcement in the dense DRA and lean DRA groups, respectively (i.e., in the dense DRA group, a mean of 21.81 alternative responses per session produced a mean of 11.72 reinforcer deliveries; in the lean DRA group, a mean of 20.81 alternative responses per session produced a mean of 5.28 reinforcer deliveries). On the other hand, previous research evaluating the effects of alternative reinforcement rate used much leaner schedules of reinforcement across groups. For example, in Craig and Shahan (2016), 4.05% and 4.23% of alternative responses produced reinforcement in their dense alternative reinforcement conditions (i.e., 84.05 and 80.58 responses per min produced 3.40 and 3.41 alternative reinforcer deliveries per min, respectively) and 2.24% and 1.92% of alternative responses produced reinforcement in their lean alternative reinforcement conditions (i.e., 40.22 and 47.28 responses per min produced .90 and .91 reinforcer deliveries per min, respectively). Other studies with nonhuman animals have used similarly lean schedules (e.g., 3.67% and 4.62% of alternative responses produced reinforcement in the low- and high-rate conditions in Craig et al., 2016, respectively; 4.08% and 13.03% of alternative responses deliveries produced reinforcement in the lean and rich conditions in Sweeney & Shahan, 2013, respectively). There may be difficulties comparing findings from experiments using such disparate schedules of reinforcement. For example, an organism’s ability to detect changes in contingencies may be different when there are larger shifts in reinforcement (e.g., in our experiment, shifts from 53.74% and 25.37% of responses producing reinforcement to 0% of responses producing reinforcement in the dense DRA and lean DRA groups, respectively) than with smaller shifts (e.g., about 4% and 2% of responses producing reinforcement to 0% of responses producing reinforcement in studies with nonhuman animals) during the transitions between phases. This procedural difference could be responsible for our divergent findings because an organism’s ability to detect changes in contingencies plays an important role in resurgence (Shahan et al., 2020).

An additional limitation in this experiment was that responding did not extinguish during Phase 3 in any experimental group. This limitation with human-operant research involving extinction has been noted elsewhere (e.g., Saini et al., 2021), and future research should focus on identifying procedural modifications to improve the efficacy of extinction-based procedures in human-operant research.

Despite these limitations, we found no evidence to suggest that on/off DRA cycling would have deleterious effects relative to the other two groups had the present study been a more applied evaluation. The top panel of Figure 1 shows that, although target responding for the cycling DRA group tended to occur at higher rates when alternative reinforcement was unavailable than when it was available, these increases in target responding rarely exceeded target response rates observed in other groups. Additionally, the top panel of Figure 3 shows that these fluctuations never caused cumulative target responding in the cycling DRA group to exceed cumulative target responding in the dense DRA or lean DRA groups. Further, our findings tended to support the effects of on/off DRA cycling identified by Shahan et al. (2020) but with human participants and using more clinically aligned procedures. Together, these data suggest future researchers should continue to evaluate the effects of on/off DRA cycling procedures in other translational and potentially applied contexts.