Influence of Televisibility and Harm Probability on Clinical-Ethical Decision Making

Abstract

Researchers have recently begun to use a behavioral economics framework to study the clinical-ethical decisions made by practicing behavior analysts. Much of this work, however, has examined broad patterns as opposed to isolating the underlying behavioral processes. In this study, we sought to extend past research by studying how clinical-ethical decisions would be influenced by a parametric manipulation of the probability that each available option would be televisible or cause short-term harm to the client. Behavior analysts (n = 15) were largely influenced only by the probability of short-term harm. In contrast, the control group (n = 30) was influenced by the probability each choice was televisible and the probability of short-term harm. Further, across all choices, control group participants showed a higher tendency than behavior analysts to not allow the individual to engage in the harmful behavior. Quantitative models built using machine learning algorithms were able to predict ~75% of choices made by participants using only the independent variables manipulated in this study. At the individual level, a probability loss discounting framework seemed to account for the data; however, deviations from traditional probability loss discounting methods provide many areas for future research. In total, the present experiment highlights the potential behavioral processes involved in clinical-ethical choices, similarities between individual (moral) and group-level (ethical) responding, and areas where practicing behavior analysts may have preferences that differ from their clients or their clients’ caregivers.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40617-023-00824-x.

Behavioral economics is a broad term that typically refers to research that combines research and theoretical approaches from the domains of economics and behavior science. At one end of the behavioral economics spectrum, researchers have sought to understand how they might bring an understanding of human behavior to bear on their home discipline of economics (e.g., Samuelson, 1937; Halpern et al., 2013; Patel et al., 2014). Research here tends to focus on the economic behavior of groups of people (e.g., purchase patterns, public health, policy decisions) and how humans’ tendency to deviate from optimal patterns of choice might be problematic or used to design better policies and large-scale interventions (e.g., List et al., 2018; Lunn & Choisdealbha, 2018). Popular examples of work through this lens include Nudge (Thaler & Sunstein, 2009), Thinking Fast and Slow (Kahneman, 2013), and form the basis of The Behavioural Insights Team founded within, though now independent of, the UK government (https://www.gov.uk/government/organisations/behavioural-insights-team).

The other end of the behavioral economics spectrum begins with behavior scientists. Here, researchers convergently evolved1 the notion of valuation wherein they assume that organisms emit choices based on momentary valuation of available outcomes as determined by operant and respondent learning (e.g., Madden, 2000; Rachlin, 1995). Here, research has focused primarily on what variables will influence the value of various outcomes and precisely how those variables influence valuation and subsequent choice at the individual level. For excellent reviews of the behavior science end of the behavioral economics spectrum, readers are referred to Gilroy et al. (2018) and Hursh and Roma (2016).

Researchers have begun to extend behavioral economics approaches to study and understand ethical decision making in clinical contexts. For example, a common finding in behavioral economics research is the “amount effect” (a.k.a. the “magnitude effect”; Rachlin et al., 2015). Here, researchers have consistently observed two patterns of human behavior. First, preference for a larger–later or larger–uncertain outcome increases when researchers increase the amount of putative reinforcers (gains) associated with two alternatives while maintaining the relative ratio between the two reinforcers (e.g., Estle et al., 2006; Myerson et al., 2011). Second, preference remains unchanged when researchers increase the amount of putative punishers (losses) associated with two alternatives while maintaining the relative ratio between the two putative punishers (e.g., Estle et al., 2006).

At the group level, researchers have observed this same pattern with ethical decisions relevant to decisions made by behavior analysts in practice. For example, across two of the scenarios in Experiment 2 of Cox (2021), the researcher asked participants to choose between two options where the amount of benefit a participant would gain from one option systematically decreased (“mix-and-match different services”) or where the amount of harm systematically increased (“report billing fraud”). Like past work on the amount effect in behavioral economics, systematically changing the amount gained from mixing-and-matching services led to systematic changes in participant choice whereas systematically changing the amount of loss from reporting billing fraud did not influence participant choice.

As a second example, past behavioral economics researchers have studied a pattern of behavior tacted as “demand” (e.g., Hursh & Silberberg, 2008; Madden, 2000). Here, researchers have consistently observed that increasing the effort required to contact a reinforcer will lead to an initial increase in responding to maintain the amount of reinforcement contacted. However, at some point, increasing the amount of effort is associated with a systematic reduction in the amount of responding emitted to contact the reinforcer. Past researchers studying ethical decision making from a behavioral economics framework have observed a potentially similar pattern where increasing the amount of extra training needed to get a caregiver to collect accurate data (i.e., effort) systematically reduced the number of behavior analysts willing to provide that extra training (Cox, 2021; Exp. 2).

One limitation to past research is that the researchers have used practicing behavior analysts or students pursuing a career in behavior analysis. Thus, it is unknown the extent to which the patterns of choice observed in previous research are specific to behavior analysts in training versus those who have been practicing independently for some time. Comparing choice between behavior analysts with greater experience and nonbehavior-analyst control groups may carry more face-validity regarding the types of interactions more likely to occur in practice. In particular, such comparisons may help practicing behavior analysts understand whether the approaches and decisions they make coincide or conflict with the choices their clients or their client’s caregivers might make in the same situation. Such information could allow practicing behavior analysts to proactively prepare for such preference differences rather than reactively responding to preference differences at the moment the caregiver voices their disagreement.

A second limitation is past research has focused primarily on group level responding. It can be argued that it makes sense that ethical behavior has been examined at this level because ethics is a group-level phenomena (as opposed to morality, which can be distinguished as an individual level phenomenon; Cox, 2020a). However, from a behavioral process perspective, group-level analyses are a less convincing demonstration of the functional control of ethical decision making that necessarily must occur at the individual level (e.g., Baum, 2005; Skinner, 1953, 1971). Thus, the current study was designed to examine ethical decision making in a manner allowing for more direct analysis of within-subject changes in choice.

A third limitation to past research examining clinical-ethical decision making is a lack of tying the examined choices to specific behavioral processes. That is, even though the experiments described above were labeled as similar to common research in behavioral economics, they were not designed to test those behavioral processes directly. As a result, the extent to which the above behavioral processes directly influenced individual choice in those scenarios is unknown. Thus, this study was designed to examine ethical decision making in behavior analysts with a more direct influence of two robust patterns of choice from the behavioral economics research literature: probability discounting and loss discounting.

Probability discounting can be defined as the systematic devaluation of an outcome as the probability decreases that the organism will contact that outcome (e.g., McKerchar & Renda, 2012; Rachlin, 2006). Loss discounting can involve probabilistic outcomes, delayed outcomes (e.g., Mazur, 1987; Rachlin, 2006), outcomes with varying effort to contact (e.g., Apps et al., 2016; Bialaszek et al., 2017), or outcomes wherein the individual contacting them are of varying social relatedness (e.g., Jones & Rachlin, 2006; Schwenke et al., 2022). What is common to all loss discounting studies, however, is that the outcomes being considered in a choice scenario involve losses of some kind (i.e., a putative punisher) as opposed to gaining something (i.e., a putative reinforcer).

Probability discounting was a logical behavioral process to test because ethical contingencies are primarily social mediated contingencies. For ethical contingencies to be contacted, someone other than the behavior analyst emitting the unethical behavior observes the unethical response and then initiates implementation of the contingency (e.g., report the incident to the BACB, report to supervisor). At the time this study was being conceptualized, this “probability of one’s clinical decisions being observed by others” was being discussed in popular social media channels under the name of “televisibility” (e.g., Animal Training Academy, 2020; Hanley, 2021; Kemmerer, 2021). For this study, we defined televisibility specifically as the probability that the person’s choice was observed by a social community and via social media (as opposed to television). This allowed us to tie televisibility more directly to probability discounting and ethical contingencies which require a larger social group.

Loss discounting was also a logical behavioral process to test as ethical decisions might be followed by outcomes involving losses. For example, the contingencies around unethical behavior emitted by certified behavior analysts might involve being reported to the Behavior Analyst Certification Board (BACB) and contacting a range of potential outcomes aimed at reducing unethical contact in the future such as mandatory supervision, revocation of certification, practice restriction, or suspension (BACB, 2023). Each of these might involve losses of resources such as time or money and future earnings by losing the ability to practice. Further, a common reason a response might be considered unethical is because the behavior results in some kind of harm to a client or the field of ABA. These outcomes also could be considered a loss within a utilitarian ethical framework (e.g., Brodhead et al., 2022).

To test the validity of probability and loss discounting influences on clinical-ethical decision making requires identification of a clinical-ethical decision that realistically would include consequences with these features. Our selection of a realistic scenario for this study came from two previous publications. The first were the two scenarios from Experiment 2 of Cox (2021) that involved tradeoffs between benefit versus harm and between client versus treatment team preferences. These features were chosen to manipulate because they led to the greatest change in responding at the group level suggesting targeting these again might lead to detectable changes in behavior at the individual level. The second publication that influenced the present study was Bannerman et al. (1990), which discusses balancing the right to habilitation with clients’ right to personal liberties. Within the article, one of the scenarios discussed is eating a preferred food that might impede overall health improvement or cause harm if engaged in regularly (e.g., eating too many doughnuts). Based on the above, we chose to test the influence probability and loss discounting for clinical-ethical scenarios where a client wants to eat a large amount of unhealthy food for a meal.

To summarize, there were three primary research questions. First, do probability and loss discounting frameworks lead to patterns of responding with ethical decisions as has been observed with past discounting research with money and other commodities? Second, given ethical behavior is a group-level phenomenon and behavior analysts are often one of many stakeholders in clinical teams, do ethical decisions made by certified behavior analysts differ than ethical decisions made by nonclinical, nonbehavior analysts? Lastly, though ethical codes are specific to groups, ethical contingencies must operate at the individual level. Our final question was how well do individual patterns of choice align with group trends and, relatedly, how well can we describe and predict choice at the individual level?

Method

Participants

Forty-five participants were recruited from either one of two modalities (see Table 1 for demographics data). Behavior analyst participants (n = 15) were recruited through behavior analytic listservs, LinkedIn, and word of mouth in organizations that provide clinical services for individuals with autism (see Table 2 for behavior analytic specific information). This involved sharing an IRB approved brief description of the study, eligibility criteria, and link to the experimental tasks in Qualtrics. Potential participants who were interested in participating could then follow the link if they opted to participate. The average (median) time to complete the experiment for the behavior analyst participants was 10.98 min (7.65 min). There were no inclusion criteria for behavior analysts other than that they self-reported as a certified behavior analyst.

Table 1.

Demographic Data for All Participants (N = 45)

Category	Behavior Analysts (n = 15)	Control Group (n = 30)
Age in Years
Average	38.86	32.77
(Median, Std. Dev.)	(36.00, 9.63)	(31.50, 10.82)
Race
White	87% (n = 13)	73% (n = 22)
Black or African American	7% (n = 1)	30% (n = 9)
Other	7% (n = 1)	3% (n = 1)
Gender
Percentage Male	33% (n = 5)	33% (n = 10)
Percentage Female	67% (n = 10)	67% (n = 20)

Note. The groups did not differ significantly in any category using a simple normal test for proportions with alpha set to 0.05.

Table 2.

Experience, Work Setting, and Populations Served for the Behavior Analyst

	Number (Percentage) of Behavior Analysts Who Selected
Setting	Clinic	Home	School	Residential	Other
	5 (33%)	4 (27%)	6 (40%)	1 (7%)	2 (13%)
Client Age (Years)	0–5	6–10	11–20	20+	Other
	9 (60%)	9 (60%)	7 (47%)	6 (40%)	1 (7%)
Population	Autism Spectrum Disorders	Disabilities Non-ASD		Other
	12 (80%)	5 (33%)		5 (33%)

Note. Participants could select more than one option so percentage sums can be greater than 100%.

Control participants (n = 30) were recruited through the platform Prolific (www.prolific.co). Prolific is an online platform where researchers can post research tasks and experiments, and people throughout the world can sign up to work as a participant in exchange for monetary compensation. Participants were recruited for a study titled, “Making Decisions About Others that People Might Find Out” with the following description:

In this study, you will be presented with a series of scenarios where you have to decide which of two clinical decisions you would make. There is no right or wrong answer. We are simply interested in how different characteristics of the situation impact what choice you make.

Potential participants could sign up to participate if they met the inclusion criteria of their first language being English and they had a minimum approval rating of 90% based on past work examining relations between approval ratings on Prolific and dishonest responding (Schild et al., 2021). The average (median) time to complete the experiment was 15.74 min (14.1 min) and participants were compensated at a base rate of $2.50, which equals an hourly rate of $10.00.

Decision Tasks

All participants answered 81 questions structured as shown in Fig. 1. For all choice scenarios, it was communicated that a client would like to eat a large amount of unhealthy food for a meal and the participant could make one of two choices: (1) allow the client to eat the meal or (2) not allow the client to eat the meal. For each question, allowing the client to eat the meal was associated with a varying probability of being televisible (defined below) and a varying probability of causing short-term harm to the client (defined below). Not allowing the client to eat the meal was associated with an independently varying probability of being televisible and an independently varying probability of short-term harm to the client.

Fig. 1.

Visual display of one of the choice scenarios presented in the experiment. All text remained the same across the 81 scenarios with the exception of the bolded text

At the beginning of each of the three blocks, participants were told (or reminded) of the definitions of short-term harm and televisibility for the current study. Short-term harm was defined as any physical or physiological discomfort within 24 hr. Televisibility was defined as the extent to which the participant’s behavior might be broadcast to social media.

We systematically varied four words across each question presented to the participants (bolded words in Fig. 1). Each bolded word corresponded to the probability of each choice being televisible (i.e., “low,” “medium,” or “high” probability) or resulting in short-term harm (i.e., “low,” “medium,” or “high” probability). All possible combinations of each of the three probabilities for each of the four probabilities were presented to every participant. Questions were clustered into three blocks around the probability that allowing the client to eat the food (first option) would be televisible. One block indicated low probability of televisibility for allowing the client to eat the food, one indicated medium probability of televisibility for allowing the client to eat the food, and one indicated high probability of televisibility for allowing the client to eat the food. Qualtrics was programmed to present each of the three blocks in a random order to each participant.

Within each block, the remaining three probabilities were altered in the same sequence (see Table 3 for the order of trials within each block). On every trial, the probability of being televisible changed on every trial for not allowing the client to eat the food. Every three trials, the probability of harm was changed for not allowing the client to eat the food. And, every nine trials, the probability of harm was changed for allowing the client to eat the food. For each of the three blocks, participants always answered questions in the sequence shown in Table 3. In total, participants made 81 discrete choices involving all combinations of the three probability levels for each of the four outcomes.

Table 3.

The Sequence of Scenarios Presented to Participants within Blocks

Scenario	Allow Client to Eat Food		Don’t Allow Client to Eat Food
	Televisibility Probability	Harm Probability	Televisibility Probability	Harm Probability
1	Low	Low	Low	Low
2	Low	Low	Medium	Low
3	Low	Low	High	Low
4	Low	Low	Low	Medium
5	Low	Low	Medium	Medium
6	Low	Low	High	Medium
7	Low	Low	Low	High
8	Low	Low	Medium	High
9	Low	Low	High	High
10	Low	Medium	Low	Low
11	Low	Medium	Medium	Low
12	Low	Medium	High	Low
13	Low	Medium	Low	Medium
14	Low	Medium	Medium	Medium
15	Low	Medium	High	Medium
16	Low	Medium	Low	High
17	Low	Medium	Medium	High
18	Low	Medium	High	High
19	Low	High	Low	Low
20	Low	High	Medium	Low
21	Low	High	High	Low
22	Low	High	Low	Medium
23	Low	High	Medium	Medium
24	Low	High	High	Medium
25	Low	High	Low	High
26	Low	High	Medium	High
27	Low	High	High	High

Note. “Televisibility Probability” is “Low” here for ease of readability, but it varied across three blocks.

Additional Questions

In addition to the experimental tasks above, participants answered nine demographics questions and seven follow-up questions. Demographics questions asked about the participants’ age, self-identified gender, self-identified cultural background, experience in ABA, the settings they work within, and the populations with whom they work (see online Supplemental Material for a copy of the task). Follow-up questions were open-ended and asked about any strategies they used during the experimental task, their experience with “televisibility” as a term, their experience with clients engaging in potentially harmful behavior but that aligns with their life preferences, and if they had any additional feedback for the researchers. A full copy of the experimental tasks and questions can be found in the online Supplementary Materials.

Data Analysis2

We first conducted a two-way repeated measures analysis of variance (ANOVA) to test for the main effects of the five independent variables in this study (i.e., to answer research questions 1 and 2). One variable was the group to which the participant belonged. The second and third variables were the probability that allowing the client to eat the food would be televisible or lead to harm, respectively. The fourth and fifth variables were the probability that not allowing the client to eat the food would be televisible or lead to harm, respectively. The results of the ANOVA provide the researcher with quantitative evidence as to whether each of the independent variables influenced choice (i.e., allowed us to answer “yes” or “no” to both research questions. However, ANOVAs do not provide information on the extent to which choice was controlled by the independent variables and the role of potential confounding variables. In other words, ANOVAs do not tell you how well you can predict any one participant’s choice based on the independent variables manipulated.

To answer the third research question, we used eight different statistical or machine learning algorithms to determine how precisely we could predict each participant’s choices using the five independent variables in this study. These algorithms ranged from simple to complex and included: logistic regression (Peng et al., 2002), k-Nearest Neighbors (Pedregosa et al., 2011), Support Vector Classifier (Chang & Lin, 2001), Stochastic Gradient Classifier (Pedregosa et al., 2011), Gaussian Process Classifier (Rasmussen & Williams, 2006), Random Forest Classifier (Breiman, 2001), Adaptive Boosting Classifier (Zhu et al., 2009), and Gradient Boosting Classifier (Friedman, 2001). There were 3,645 total choices in this study. For each algorithm, we used five-fold with 10-repeat cross validation and loss metrics calculated using a 20% holdout test set. That is, we fit a model on a random selection of 2,916 of the observations (80%), used that model to make predictions on the remaining 729 observations (20%), and saved the loss metrics associated with each model for analysis.

Finally, we were also interested in how each individual participant’s choices were influenced by the four probabilities manipulated in this study. To analyze choice at the individual level, we used the top performing modeling approach from above (logistic regression) to build one model for each of the 45 participants using data from their 81 choices. These analyses subsequently allowed us to compare the influence of each of the four probabilities manipulated on participant choice at the individual level.

All quantitative analyses were conducted using the statsmodels (Seabold & Perktold, 2010), pingouin (Vallat, 2018), scipy (Virtanen et al., 2020), or scikit-learn (Pedregosa et al., 2011) package for Python (version 3.9.13). The data and code used for the analyses can be found at Cox and Javed (2022).

Results

Main Effects

Figure 2 shows bar plots of the percentage of choices the participants did not allow the client to eat the food (y-axis) based on the probability (x-axes) that the participant’s choice was televisible (left panels) or would result in short-term harm (right panels). Error bars represent the 95% confidence interval.3 Overall, we observed significant differences in participants’ not allowing the client to eat the unhealthy food as a function of the experimental group (F_{(1, 43)} = 4.16; p = .0475; 𝜂² = 0.0883) and the unique combination of four probabilities across both choices (F_{(23, 989)} = 32.68; p < .001; 𝜂² = 0.4318). This is observed visually as well. Because the probability increased that allowing the client to eat the unhealthy food was televisible and caused harm, we observed a systematic increase in the percentage of participants who chose to not allow the client to eat the unhealthy food. Likewise, because the probability increased that not allowing the client to eat the unhealthy food was televisible and caused harm, we observed a systematic reduction in the percentage of participants who did not allow the client to eat the unhealthy food. As a generalized observation, as the probability of televisibility and harm increased for one behavior, participants increasingly emitted the alternative behavior. However, for both choices the probability of harm influenced behavior more than the probability the choice was televisible.

Fig. 2.

Overall influence of the probability of televisibility and harm on all 45 participants’ choice to “Allow” or “Don’t Allow” the client to eat the preferred food item. Error bars represent 95% confidence intervals

Group Differences

Figure 3 shows the same bar plots as Fig. 2 except separated by group. The black bars represent the average percentage of choices made by behavior analysts and the gray bars represent the average percentage of choices made by the control group. Across all four independent variables manipulated in this study, the control group participants were much more likely to not allow the client to eat the large amount of unhealthy food regardless of the probability that the choice was televisible or the probability that it would lead to harm.

Fig. 3.

Overall influence of televisibility, harm, and respondent background on participant decision to “Allow” or “Don’t Allow” the client to eat the preferred food item. Error bars represent 95% confidence intervals

For behavior analysts, changing the televisibility probability for allowing the client to eat the unhealthy food and for not allowing the client to eat the unhealthy food had no influence on choice (left two panels; Allow: F_{(2, 28}) = 0.34; p = .71; 𝜂² = 0.0013; Don’t Allow: F_{(2, 28)} = 0.49; p = .62; 𝜂² = 0.0009). However, changing the probability that either choice would cause harm systematically influenced choice in the directions described above. Increasing the probability that allowing the client to eat food would cause harm increased the percentage of times these behavior analysts chose to not allow the client to eat the food (upper right panel; F_{(2, 28)} = 16.67; p < .001; 𝜂² = 0.2004). And increasing the probability that not allowing the client to eat food would cause harm decreased the percentage of times they chose not to allow the client to eat the unhealthy food (lower right panel; F_{(2, 28)} = 5.10; p = .0129; 𝜂² = 0.0676).

For the control group, increasing the probability that allowing the client to eat the food was televisible increased the percentage of participants who chose not to allow the client to eat the unhealthy food (upper left panel; F_{(2, 58)} = 10.44; p = .0001; 𝜂² = 0.0479). Likewise, increasing the probability that not allowing the client to eat the food was televisible decreased the percentage of participants who chose not to allow the client to eat the unhealthy food (lower left panel; F_{(2, 58)} = 7.31; p = .0015; 𝜂² = 0.2014). Likewise, changing the probability that either choice would cause harm systematically influenced choice for the control group in the directions described above. Increasing the probability that allowing the client to eat food would cause harm increased the percentage of times the control group participants chose to not allow the client to eat the food (F_{(2, 58)} = 43.73; p < .001; 𝜂² = 0.6013). And increasing the probability that not allowing the client to eat food would cause harm decreased the percentage of times they chose not to allow the client to eat the unhealthy food (F_{(2, 58)} = 15.47; p = .0003; 𝜂² = 0.3479).

Group Modeling

Figure 4 shows the fit metrics for the eight different statistical or machine learning algorithms aimed at predicting each response made by participants using the five independent variables in the present study: (1) group; (2) harm probability for allowing the client to eat the food; (3) harm probability for not allowing the client to eat the food; (4) televisibility probability for allowing the client to eat the food; and (5) televisibility probability for not allowing the client to eat the food. For each metric, a value of 1.0 would indicate the model correctly predicts every choice made by a participant and a value of 0.0 would indicate the model gets zero predictions correct. Thus, for each metric, values closer to 1.0 indicate the model more closely predicts participant choices.4

Fig. 4.

Results of multi-model testing to determine how we can predict which choice a participant would make using the probability of televisibility, probability of harm, and group variables for each participant

The algorithm with the best average predictive accuracy (71.35%) and best average precision (75.17%) was the Gradient Boosting Classifier. The algorithm with the best average recall (79.32%), F1 (75.74%), and Receiver Operating Characteristic Area Under the Curve (ROC-AUC; 76.34%) was logistic regression. However, as can be observed visually, all algorithms were relatively similar in their ability to develop models that could predict whether a participant would choose to allow or not allow the client to eat the food based on the televisibility probabilities, harm probabilities, and learning history with behavior analysis. Further, with the exception of the Gaussian Process Classifier, all algorithms led to models with relatively stable predictive accuracy (all ranges within 6.7%), precision (all ranges within 12.2%), recall (all ranges within 10.3%), F1 (all ranges within 7.4%), and ROC-AUC (all ranges within 8.0%). In sum, at the group level, approximately three out of every four of choices made by a participant could be predicted using the five predictor variables manipulated in this study.

Individual Decision-Making

We also examined how each individual participant’s choices changed as a function of televisibility and harm probability. Figure 5 shows data from four participants whose data are exemplary of the four patterns of responses we observed. Nine participants always chose the same option, regardless of the probabilities associated with either option (upper left panel). Four of those nine participants always chose to allow the client to eat the food (three behavior analysts; one individual from the control group) and five of those nine participants always chose to not allow the client to eat the food (two behavior analysts; three from the control group).

Fig. 5.

Example data from four participants that show trends in the influence of harm and televisibility on decisions at the individual level

The remaining 36 participants were influenced in some capacity by the four probabilities manipulated in this study. Nineteen of the participants were only influenced by the probability of harm associated with each option (upper right panel; 8 behavior analysts, 11 from the control group). It can be argued that one participant from the control group was only influenced by the televisibility of allowing the client to eat the food (lower left panel)—though that was primarily due to a difference between “low” and “medium” televisibility probability when allowing the client to eat the food. Lastly, 16 participants were influenced by both the probability of their choice being televisible and causing harm (lower right panel; 2 behavior analysts, 14 from the control group). In sum, 35 of the 45 participants were influenced by the probability of harm associated with each option and 16 participants were also influenced by the probability that their choice would be televisible.

Figure 6 shows the results of using the best performing algorithm to build individual logistic regression models for each participant. The average (median, standard deviation) for model accuracy was 88.81% (93.83%, 12.62%) and model accuracy was greater than 85% for 36 of the participants (far left panel). This suggests that knowing the individual participant’s learning history with behavior analysis as well as the four probability levels allowed us to predict most of the individual participants’ choices between the two available options.

Fig. 6.

Average accuracy and resulting beta weights for each logistic regression model that was fit to an individual participants’ choices. Each marker represents model accuracy or the estimated beta weights for one participant

The second-from-left and middle panel in Fig. 6 show the estimated beta weights from each model fit for each participant relative to televisibility. Beta weights quantify the relative influence that independent variable has on choice for that participant. Values further from 0.0 indicate a greater influence on choice. Negative values indicate an increased likelihood that the participant would choose the other option and positive values indicate an increased likelihood the participant would choose that option. The average (median, standard deviation) beta weight for the probability of televisibility for allowing the client to eat the food was 0.37 (0.40, 0.63) and the average (median, standard deviation) beta weight for the probability of televisibility for not allowing the client to eat the food was -0.21 (-0.19, 0.54).

The second-from-right and far right panels in Fig. 6 show the estimated beta weights from each model fit for each participant relative to harm. The average (median, standard deviation) beta weights for the probability of harm for allowing the client to eat the food was 1.61 (1.94, 0.99) and the average (median, standard deviation) beta weights for the probability of harm for not allowing the client to eat the food was -1.02 (-1.60, 1.31).

Discussion

We examined how people changed their choice to allow a client to eat a large amount of unhealthy food for a meal as a function of the probability that the decision would be televisible or lead to short-term harm for a client. There were three main findings in response to the three primary research questions. First, the probability that a response would lead to harm influenced choice to a greater extent than the probability the choice was televisible. Second, behavior analysts and the control group differed significantly in the extent to which all four probabilities influenced choice. Finally, patterns at the individual level largely mirrored those at the group level and approximately 8-to-9 out of every 10 choices made could be predicted using logistic regression models of the independent variables manipulated. We discuss each of these main findings in turn.

Influence of Harm and Televisibility on Choice

Seventy-eight percent of participants’ choices in this study were influenced by the probability that the choice would lead to short-term harm. In particular, because the probability of harm increased for emitting one response (either allow the client to eat the food or don’t allow the client to eat the food), participants were more likely to emit the alternative response. This pattern coincides with observations made by previous researchers studying probability loss discounting (e.g., Estle et al., 2006). That is, as the probability increases that one response will contact a loss or punisher, organisms tend to systematically reallocate their responding to concurrently available alternatives. Given the patterns of choice observed in this study, it seems likely that the same behavioral processes involved in probability loss discounting also were involved in choice in the present study.

Probability loss discounting studies typically use an explicit amount of punisher (e.g., losing $100) and it is implied that the contingency will be implemented with perfect fidelity. This study differed along both dimensions. We defined short-term harm as any physical or physiological discomfort within 24 hr. This is much vaguer than most probability discounting studies. Though the time frame was explicit, different participants are likely to have different histories with “discomfort,” which may have influenced how they responded in the present experiment. We took the approach we did in this study for ease of scenario presentation and because this was a first attempt at this kind of empirical setup. But this difference suggests an interesting direction for future research relative to how individual histories with risk-benefit framing might influence clinical-ethical decisions.

Past work around framing effects suggests the specific language used when presenting choices to participants will influence choice (e.g., Bonner et al., 2021; Yi & Bickel, 2005). Thus, we suspect much future research could examine how framing ethical outcomes influences choice. Exactly how the different options might be framed in future research are legion and have come to be labeled by some researchers with the broad classification scheme of “choice architecture” (e.g., Münscher et al., 2015). Examples include whether the exact same options are framed as gains or losses, whether default or zero valued options are explicit, whether probabilities are described using percentages or frequencies, and the degree of specificity or abstractness used in describing outcomes.

The second way this study differed from past probability loss discounting studies are relative to the probabilities of contingencies being implemented. Many clinical decisions (such as the one in this study) are likely to be observed only by the client and the handful of individuals who are part of the intervention team. But ethical contingencies may require a greater number of people in a social community to be implemented (e.g., a colleague knowledgeable with the BACB Ethics Code and who is likely to report violations). The probability that one’s clinical decisions would be broadcast to a larger social community (e.g., televisibility) increases the likelihood that ethical contingencies might be implemented. This might be conceptually similar to treatment fidelity studies conducted by past researchers where the probability that contingencies are implemented have been systematically manipulated with a resulting influence on the behaviors that contact the contingencies (e.g., St. Peter et al., 2016). Thirty-eight percent of the participants in the current study were influenced by the probability their choice would be televisible and in a direction similar to past research on treatment fidelity. However, this study was designed to examine televisibility in general rather than the fidelity of ethical contingencies in applied practice. As with the discomfort manipulation above, this leaves open the door for many possible directions for future research on basic processes and parallels to other areas of the behavior analytic literature.

The analysis of more complex choice scenarios is one final and important contribution of the present study to the probability loss discounting literature. Many situations in life (clinical or otherwise) involve choice between many different alternatives, each associated with different delays, probabilities, and effort; and where each alternative likely contacts multiple outcomes (e.g., Cox & Dallery, 2018). In short, many clinical-ethical decisions involve concurrent choice with multidimensional outcomes. Despite the commonality of these complex choice scenarios, most studies on choice have analyzed one or two dimensions and for a single outcome.

Examining concurrent choice arrangements within multidimensional contexts is useful for behavior analysts in practice. This is because researchers can begin to determine whether one variable influences choice more than others. In the current study, examining multidimensional choice across televisibility and harm allowed us to determine that the potential harm to a client was much more influential than the probability the choice would be observed on social media. In turn, ethics coordinators (Brodhead & Higbee, 2012) or ethics committees (Cox, 2020b; LeBlanc et al., 2020) can focus their efforts and attention on contingencies likely to have the most impact on the decisions that practitioners make in the organizations that they serve. Here, future research directions could span the basic-to-applied continuum. From a basic research standpoint, more variables could be examined and paired with one another to identify how they interact to influence decision making. From an applied perspective, important steps include developing efficient methods to assess ethical preference in practitioners and clients/caregivers and how to make these results actionable and efficiently fit in existing organizational systems.

Behavior Analysts versus General Population

The second set of primary findings surrounded the difference in choice between behavior analysts and the control group. The first main difference was that a statistically and visually greater percentage of control group participants chose to not allow the client to eat the food as compared to behavior analysts. Ethics are defined relative to a larger social group that share a common profession (e.g., board certified behavior analysts) or learning history related to a “belief system” (e.g., religion, political ideology). The study here highlights how individuals who are not a member of the social group tacted as “behavior analysts” may have preferences in clinical-ethical decision-making contexts that differ in some way from behavior analysts. For the specific situations examined in this study, the data suggest that parents or caregivers might be less tolerable in general than behavior analysts in allowing clients to engage in behavior that leads to short-term harm. It also highlights how behavior analysts may benefit from proactive conversations on preference alignment with caregivers around these topics rather than reactive conversations once a misaligned decision has been made.

No known task exists for efficiently measuring the ethical preferences of practitioners and the clients they serve (or their caregivers). If one were created, however, one could envision regular assessment of practitioner and client (or their caregiver) ethical preferences to proactively know where alignment and misalignment occurs. Past research suggests values disagreement is a primary challenge to clinical ethics resolution and consensus rarely follows once a disagreement arises (e.g., Bell et al., 2022; Gornick & Zikmund-Fisher, 2019). Future research here might involve (1) identifying ways to assess ethical preference proactively; (2) understanding how preference influences clinical-ethical decision making contexts; and (3) how to proactively navigate conflict among practitioners and clients (or their caregivers) where it exists.

A second difference between behavior analysts and the control group was the influence of televisibility on choice. Only 13% of the behavior analysts reported that the probability of their choice being televisible would influence their decisions. In contrast, 47% of the control group participants reported their choice was influenced by the probability it was televisible. The sample sizes used in the present study disallow broad claims about all behavior analysts or all nonbehavior analysts. However, a general observation is noteworthy in that there might be variables relevant to a clinical-ethical decisions that have no influence on a behavior analyst’s decision and they therefore do not consider it when making a clinical-ethical decision. But that same variable might be important to the preferences of the client or their caregiver. If these variables are “unknown-unknowns” (Rumsfeld, 2002) to behavior analysts, decisions where these variables are present may result in conflict between behavior analysts and their clients or caregivers. Future research that continues to identify variables important to behavior analysts but not clients and vice versa will likely help behavior analysts proactively seek out and include client preferences in a greater proportion of clinical decisions rather than including them reactively after a misstep has been made.

Modeling Individual Decision Making

A final set of primary findings involves the interplay between decision making at the group and individual level. Overall, 36 of the 45 participants (80%) showed some pattern of choice that coincided with the patterns observed at the group level. This is consistent with much published research in behavior analysis. For example, many behavior analytic studies in the Journal of Applied Behavior Analysis and the Journal of the Experimental Analysis of Behavior demonstrate an effect in four out of five participants (80%), five out of six participants (83%), or similar proportions. From a behavioral process perspective, this suggests that the current study is likely identifying and isolating the variables that allow for the description, prediction, and control of clinical-ethical decisions like those in the present study. It also highlights that there is still much work to be done (in this study and others) to isolate the additional variables needed to account for choices made by the remaining 20% of participants.

One limitation to the present study was the specific clinical-ethical scenario to which participants responded (i.e., balancing short-term harm with a client’s preference to eat a preferred food). We chose this situation because it allowed a first pass at the primary components of the clinical-ethical decision context we were interested in studying: a client’s preference causes them harm, and how does the probability of that harm and the likelihood of the participant’s decision being observed socially influence that decision? We chose a large, unhealthy meal because existing literature had already analyzed and discussed the nuances of these kinds of clinical-ethical decisions (Bannerman et al., 1990). However, a common finding in behavioral economics research is that the commodity under consideration will influence choice (e.g., Chapman & Elstein, 1995; Cox et al., 2020; Weatherly et al., 2010). Thus, we suspect changing the specific client preferred behavior under consideration will also likely influence choice. However, this is an empirical question open to future researchers.

These data also fit with a growing body of research finding that different behavior analysts who are in identical ethical or clinical decision contexts may make different choices (e.g., Cox, 2021; Cox & Brodhead, 2021; Saini et al., 2018; Vanselow et al., 2011). The benefit of the current study related to past work in ethics is the use of behavior analysts with more years of experience and postcertification (min, median, avg, max = 3, 10, 12.8, 30 reported years of experience). This suggests that greater experience and training in behavior analysis does not decrease variability in ethical decisions observed among behavior analysts. Nevertheless, with only 15 behavior analysts in the present study, analysis of how years of experience covary with patterns of clinical-ethical decision-making were not possible without making many assumptions. As a result, we chose to forego these analyses. Future research interested in how years of experience influence clinical-ethical decision-making will likely need to recruit a much larger sample to account for the many covariates likely to be present.

A fundamental question is whether it matters that different behavior analysts make different ethical decisions when put in the exact same context. Is variability in ethical decisions something to celebrate or to be concerned with? Our opinion is that variability in clinical-ethical decision making is likely beneficial to the field in the long run. We believe this because there are unlikely to be any clinical-ethical rules or approaches that are optimal in all situations. Differences in clinical-ethical decision making introduce variability. If social contingencies never changed, variable clinical-ethical decision making might be problematic as variability would likely fail to maximize reinforcement and minimize punishment (e.g., Gonzalez & Dutt, 2016). But, alas, social and ethical contingencies are always changing (see human rights changes during the 20^th century). When contingencies change, response variability is needed to ensure that responding contacts the changed social reinforcement contingencies (e.g., Gonzalez & Dutt, 2016; Krueger et al., 2017). Critical questions here become: (1) whether specific clinical-ethical decisions are optimal within specific decision contexts; (2) what variables increase exploration and variability in clinical-ethical decisions across decision contexts; and (3) how ABA—as a field—can create the right amount of variability to adapt to changing social contingencies while maximizing ethical behavior in the short-term.

Finally, central to the previous paragraph is another fundamental question to the functional analysis of ethical behavior. That is, how exactly do individual ethical contingencies aggregate into professional ethical codes of ethics and rules that dynamically change over time? And how do you practically shift the ethical behavior of a group of people when you are unable to observe, measure, and manipulate all instances of their behavior or the controlling contingencies? Many philosophers throughout history have offered their opinion on the former topic. And researchers working in behavioral economics from the economics and public health side have offered answers for the latter. Future research just might bridge that gap.

Data Availability

Data and code for the analyses are available at a publicly accessible repository: https://zenodo.org/badge/latestdoi/532253440

Declarations

Conflict of Interest

None

Ethical Statement

All procedures were approved by the Endicott College IRB #1820460-2. All participants completed the IRB approved informed consent process.

RethinkFirst was not involved in any way with the research conducted, the decisions made, or the analytic strategies used in connection with this article. The opinions expressed in this article are the authors’ own and do not necessarily represent the views of RethinkFirst.