Pattern-setting for improving risky decision-making

Self-control can be defined as making choices in accordance with long-term, rather than short-term, patterns of behavior. Thinking on these lines, Howard Rachlin suggested a novel technique to enhance self-control, by making individual choices carry the weight of a larger pattern of choices. This brief report describes a study of 169 college students who made repeated choices between two gambles. One gamble dominated the other, but required waiting an uncertain amount of time. Some subjects were forced to repeat their previous choices according to a schedule. I found that such subjects chose the better gamble more often (by about 10 percentage points) than subjects who could always choose freely or who were forced to choose the better gamble. However, this effect was visible immediately, rather than emerging from learning over the course of the task as hypothesized. These results support the use of pattern-setting as a strategy to improve decision-making.

Acknowledgments

This study was originally conceived, designed, and analyzed in 2015 with the substantial involvement of Howard Rachlin (1935 – 2021).

Introduction

Failures of self-control are an all too familiar experience in human life, ranging from the mundane problem of eating too much of something tasty and getting a bellyache, to the potentially lethal crisis of severe drug addiction. But defining self-control, that is, defining behavior that constitutes a self-control success or a self-control failure, can be tricky. Rachlin (2017) proposes that self-control is behavior that is directed by higher-level, more temporally extensive goals, such as being healthy, in contrast to immediate goals, such as the short-term pleasure of another bite of food. This approach naturally fits into what Rachlin calls teleological behaviorism, the view that an organism's mind is best identified with the complex set of overlapping, often very long-term behavioral patterns that direct its behavior over its life. Individual moment-to-moment actions, such as walking one step, occur as part of larger goal-directed behaviors, such as going to work, and patterns telescope outwards in this fashion until we reach the organism's highest-order, most abstract goals, such as being a good person or living a fulfilling life. To act with self-control, then, is to act in accordance with a long and important pattern rather than a competing shorter and less important pattern. In fact, Rachlin (2019) argues that a behavioral-evolutionary process can explain how agents can settle on large behavioral patterns that are on net beneficial even when those patterns comprise many smaller behaviors that are individually deleterious.

If an agent's long and short behavioral patterns are in conflict, and obeying the longer patterns is to its benefit, then it may benefit from being somehow committed to obeying the longer patterns. Thus the value of commitment devices, by which people can reduce temptation or opportunity to renege on a goal (Bryan, Karlan, & Nelson, 2010; Rogers, Milkman, & Volpp, 2014). For example, Giné, Karlan, and Zinman (2010) found that when smokers were given the opportunity to put money in a savings account that they only got back if they stopped smoking, they stopped at a higher rate than controls. Although not forced to stop smoking, the smokers in the savings-account group had been given a monetary incentive to avoid this self-control failure. Ariely and Wertenbroch (2002) found that college students who had deadlines distributed throughout a several-week period, instead of having all their work due at the very end of the period, did better work, perhaps because they'd been prevented from procrastinating. Bryan et al. (2010) discusses rotating savings and credit associations (ROSCAs), which provide incentive to save money by requiring regular saving to get one's share of a communal pot.

However, monetary incentive isn't the only sort of commitment that ROSCAs provide. ROSCAs are small organizations, generally formed among people who already know each other, which leads to social pressure to continue saving. This softer kind of commitment, compared to losing money (or taking a penalty to a course grade, as in Ariely & Wertenbroch, 2002, Study 1), may still be effective. Rachlin (2016) proposes a particularly soft commitment strategy called pattern-setting (previously described in Rachlin, 2014). The idea is that once a person has identified a self-control problem, such as an addiction to cigarettes, they begin by getting into the habit of recording how often they do the undesirable behavior. Then they set up a schedule of pattern days and free days, in which on pattern days, they're obliged to do the behavior as many times as they did in the foregoing free day; for example, they must smoke the same number of cigarettes, no more and no less. Rachlin argues that by establishing large numbers of pattern days, the person will learn to associate each instance of the undesirable behavior on free days with its longer-term effect on pattern days. Thus a smoker may opt to smoke less on a free day because one cigarette is no longer just one cigarette; instead, it implies three or five or seven cigarettes smoked over the succeeding pattern days. If successful, this strategy would be a direct way to bring one's behavior into alignment with longer patterns, and thus increase self-control, despite never actually requiring or even explicitly incentivizing reduction of the undesirable behavior.

Read, Loewenstein, and Kalyanaraman (1999) examined a situation similar to pattern-setting in a study of choice for movies. People were given a list of movies they could watch, some highbrow and some lowbrow, and either selected each movie they would watch on the day they would watch it or selected three movies in advance. Read et al. found that those choosing movies in advance were more likely to choose highbrow movies, supporting the idea that choosing in advance leads to more frequent "virtuous" (larger-goal-congruent) choices. Siegel and Rachlin (1995) examined various conditions in which pigeons chose between a smaller-sooner (SS) and a larger-later (LL) food reward. When only one keypeck was required for a reward, all subjects preferred SS. When 31 pecks were required, however, the pigeons learned to prefer LL, even though only the 31st peck actually determined the reward type, meaning that a pigeon who pecked the LL key 30 times in a row could still change back to SS on the 31st trial. Thus, behaving in larger units seems to have nudged the pigeons towards the larger reward that required more patience.

In this brief report, I describe a direct test of pattern-setting as a means of improving decision-making. I used a risky decision-making task based on Luhmann, Ishida, and Hajcak (2011) (to be replicated by Ciria et al., 2021), in which human subjects could choose between two gambles, one of which dominated the other but was only available after a delay. I hoped to find that if subjects were forced to repeat their past choices in certain pattern trials, they would more readily learn to take the better gamble. Furthermore, I expected to see a weaker effect (or no such effect) among subjects simply forced to take the better gamble, instead of being forced to repeat their own choices.

Method

Task code, raw data, and data-analysis code can be found at http://arfer.net/projects/pattern.

The core task of the study was a computer-based gambling task adapted from Luhmann et al. (2011). The first 110 subjects were paid their winnings in cash, but my funds were limited, so the remaining 67 completed the task with hypothetical gambles instead.

On each trial of the task, subjects chose between two gambles, which were displayed graphically as in Figure 1. The immediate choice I (named "A" in the instructions and interface shown to subjects) had a 60% chance of paying the subject 4 cents, whereas the delayed choice D (named "B" for subjects) had a 70% chance of paying the subject 4 cents. Since D provides a higher probability of winning the same amount, it dominates I. However, D wasn't available at the start of each trial whereas I was. The wait until D could be chosen was 5 s plus an random exponentially distributed amount of time with median 5 s (or equivalently, mean $\frac{5}{\ln 2} ≈ 7.2$). The use of an exponential distribution ensured that after the initial fixed 5 s, the expected remaining wait time remained constant, so waiting provided no new information about the time remaining to wait.

When the subject chose a gamble, they were immediately told whether they'd won ("WIN! 4 cents") or lost ("0 cents"). However, if they chose I, they had to wait until the next trial. The duration of this wait was the same as the wait for D to become available, minus any time the subject had already spent waiting for D. Thus, choosing I only let the subject end the trial and see whether they'd won sooner. It didn't let them complete the whole task sooner.

Trials were grouped into 3-trial blocks. Even-numbered blocks (counting the first block as block 1) were shown with a yellow background, and odd blocks used a blue background. The significance of the blocks varied according to which of four conditions the subject was randomly assigned to. Conditions were assigned by building a sequence of randomly permuted four-condition subsequences and then assigning each subject the last unused condition. Thus I ensured that each subject was equally likely to receive each condition, while maximizing the equality of sample sizes between conditions.

• In the Control condition, subjects could choose freely on all trials.
• In the Within-Pattern condition, subjects could choose freely on trial 1 of each block, but had to choose the same gamble on trials 2 and 3 of the same block.
• In the Across-Pattern condition, subjects could choose freely in odd-numbered blocks, but had to make the same sequence of choices in each even-numbered block as they had in the previous block.
• In the Force-D condition, subjects could choose freely in odd-numbered blocks, but could only choose D in even-numbered blocks.

In each case, a reminder of the rules for the current trial were shown on screen, above the gambles. Trials on which the subject wasn't forced to choose a given gamble due to condition-specific rules are termed "free" trials henceforth.

I expected that in the two pattern conditions, compared to Control, I would see a greater rate of free choices for D in later blocks (after sufficient opportunity for learning). I expected a greater such effect in Across-Pattern, because it enforced a larger-scale pattern. Finally, I expected that free choices under Force-D would be similar to those in Control, showing that it isn't merely being forced to make choices but pattern-setting that increases free choices for D.

Procedure

Undergraduate students in the psychology department of Stony Brook University were run in September, October, and November of 2015. After providing informed consent, subjects read detailed instructions for the task (see supplement). A four-item quiz tested the subject's understanding and reinforced it by telling them the correct answer immediately after each question. Next subjects were shown their condition-specific instructions. Subjects completed 20 blocks of 3 trials. (Thus, Control subjects made a total of 60 free choices, Across-Pattern and Force-D subjects made 30, and Within-Pattern subjects made 20.) Finally, subjects were debriefed. Some subjects were compensated with cash (of an amount equal to their winnings from the task), and all received course credit.

Results

Demographics

Demographic information was obtained from screener forms the subjects filled out as part of registering for the department's subject pool. Of 177 subjects, 65% were female. Ages ranged from 17 to 49, with 95% of subjects being 22 or younger. In terms of ancestry, one item allowed the respondent to choose one of "African American/Black", "American Indian or Alaska Native", "Asian", "Caucasian/White", "Multiple Ethnicity", "Native Hawaiian or Other Pacific Islander", or "Other", and another item asked if the respondent was Hispanic. Thus 45% of subjects indicated they were white, 24% Asian, 12% black, 11% "Other", and 7.9% multi-ethnic, while 16% were Hispanic. All subjects indicated they were native speakers of English. Finally, 10% were left-handed.

Quiz

Subjects initially had some trouble understanding the task instructions. Only 53% answered all 4 quiz questions correctly, and 20% answered 2 or less correctly. Since the right answer was shown to subjects immediately after they responded, ideally correcting any misconceptions, I don't omit any subjects from analysis on the basis of quiz performance.

Gambles

While observing the gambling task, experimenters noticed 2 subjects who appeared to be inattentive: one seemed to fall asleep and one frequently checked her phone. An additional 6 subjects, on 3 or more trials, chose I after D had been available for 2 s or longer. With these 8 subjects excluded, 169 remained: 42 Control, 43 Within-Pattern, 42 Across-Pattern, and 42 Force-D. Completion times for the gambling task (not including the instructions and quiz) ranged from 13 to 23 min, with a mean of 16 min 58 s.

Table 1 shows that in the first half of trials, between 58% and 69% of free choices were for D. Such proportions, being neither near 0 nor near 1, are helpful in that they provide room to show learning effects. However, the rate of free D choices is similar in the second half. Looking at mean differences between the first and second halves (see also Figure 2), we see little evidence of change over the course of the study, or of per-condition effects on such change. The mean differences of greatest magnitude are for Control, in which subjects chose D slightly more often later in the experiment, and Within-Pattern, in which subjects chose D slightly less often later (contrary to hypothesis). Any effect by which Within-Pattern or Across-Pattern help subjects learn to choose D is small, not exceeding 6 percentage points. (Raw rate differences are in some ways a cruder measure than odds ratios or log odds ratios, but they're always real numbers even when rates of 0 or 1 are involved, and with no regression model, there's no need to attribute implausible additive probability effects to continuous predictors.)

Table 1. The proportion of free choices for D in each half of the task, across all subjects in each condition. The last column shows the mean (and 95% confidence interval) of within-subject differences between the first and second halves; positive numbers indicate more D choices in the second half than the first. Confidence intervals are computed with bias-corrected accelerated bootstrapping.
Condition 1st 2nd 2nd − 1st
Control .59 .61 +.025 [−.020, +.068]
Within-Pattern .69 .67 −.019 [−.088, +.037]
Across-Pattern .69 .69 +.006 [−.037, +.057]
Force-D .59 .59 +.006 [−.052, +.067]

Although little learning is evident, the proportions in Table 1 are suggestive of between-subject conditions that are observable throughout the task. Thus, in Table 2 (see also Figure 3), I compare proportions of free choices for D throughout the whole task. We see that the two pattern conditions produced similar rates, as did the two other conditions, but either pattern condition, compared to either of the other two, seems to have been effective in increasing choices for D, by about 10 percentage points. The final row of Table 2 lumps the two similar pairs of conditions together and then compares the pairs to each other. This analysis estimates that the pattern conditions increased choices for D by 8.8 percentage points.

Table 2. Between-condition differences in proportions of free trials for D, with 95% confidence intervals.
Conditions Difference in D choices
Across-Pattern − Control +.090 [+.058, +.121]
Across-Pattern − Force-D +.102 [+.063, +.138]
Within-Pattern − Control +.076 [+.039, +.113]
Within-Pattern − Force-D +.088 [+.046, +.129]
Across-Pattern − Within-Pattern +.014 [−.026, +.054]
Control − Force-D +.012 [−.022, +.044]
(Both pattern Cs) − (Other two Cs) +.088 [+.063, +.113]

Discussion

I found that when people were obliged to repeat choices, they were more likely to take a better option that required patience. This finding supports Rachlin's (2016) idea of pattern-setting: when their individual choices were given more weight by a pattern structure, people seemed to prioritize a larger goal (earning money, real or imaginary, from the task) over a smaller one (ending a momentary unpleasant period of waiting). And yet, this effect didn't come about from learning during the task. Regardless of condition, people chose the dominating option, D, at similar rates in the first and second half of the task: within-subject differences between the halves were small.

The fact that the between-condition differences I expected were visible immediately, rather than after a process of learning, is perhaps a testament to the self-knowledge and life experience that subjects brought to the task. People may undercorrect for their tendency to make suboptimal decisions (as in Ariely & Wertenbroch, 2002, where externally set deadlines were more effective than self-selected deadlines), but they're still savvy enough to take opportunities for correction in the first place, as by saving money that they'll only be able to get back if they stick to a plan (Bryan et al., 2010; Giné et al., 2010). So, once subjects in the pattern conditions understood the special pattern constraints they'd be subject to, they may've immediately seen the importance of not committing themselves to the worse gamble in pattern trials. A learning effect may've been more visible if the pattern constraints hadn't been explained up front, and subjects had had to learn on their own how the task worked. Such an arrangement would more closely resemble an animal study, but would also entangle what are in principle two distinct learning problems: discovering the task rules and adapting decision-making to those rules.

Considering how condition instructions seemed to suffice to produce the between-condition differences, it may be surprising how poorly subjects performed on the quiz. Only about half correctly answered all four comprehension questions about instructions they'd just read. The task program's own corrections, and the reminder of trial-specific rules shown during each trial, may've helped. A limitation is that the quiz didn't cover condition-specific instructions, and it seems obvious that not all conditions are equally easy to understand.

I expected that the larger-scale pattern condition, Between-Pattern, would have a stronger effect than the smaller one, Within-Pattern. The difference I obtained is of the right sign but close to 0, suggesting these two conditions have about the same effect. This may mean that subjects were sensitive to whether they'd be forced to repeat choices, but insensitive to the quantitative details. The fact that the control condition was similar to a condition with forced D choices (with no patterning) is helpful, because it shows the specific effect of being responsible for one's own future choices.

Overall, this laboratory test of the viability of pattern-setting as a self-control technique succeeded, albeit not in the way I expected. Would pattern-setting then be effective for real-world self-control problems, like drug addiction? Only a direct empirical test could settle this, but it's worth considering how this study's methods differ from how Rachlin (2016) proposes pattern-setting would be used in the field. First, I offered some subjects hypothetical rather than real money, and studies investigating how hypothetical rewards differ from real ones in risky choice have produced inconsistent results (e.g., Hinvest & Anderson, 2010; Barreda-Tarrazona, Jaramillo-Gutiérrez, Navarro-Martínez, & Sabater-Grande, 2011; Xu et al., 2018; Horn & Freund, 2022). Second, I didn't require subjects to monitor and record their own behavior. The task took care of this for them. Third, and perhaps more important, subjects had no choice about obeying the pattern rules. It's one thing to smoke as many cigarettes as you want on Monday and say you'll smoke the same number on Tuesday; it's another to do on Tuesday what you said you'd do, which could itself become a self-control problem. A future test of pattern-setting might try telling subjects what the pattern rule implies they should do rather than forcing them. Voluntary and consistent compliance with the pattern rule seems necessary for pattern-setting to realize its potential.

Supplement

The initial instructions and the quiz were as follows:

Some quick notes before we begin:

• Please give the experiment your undivided attention. Doing something else (like checking your phone) during a waiting period would interfere with the purpose of the experiment.
• This experiment uses timers to make you wait for certain things. Don't use your browser's back button or refresh button on a page with a timer, or the timer may restart (in which case it will have the same length as before).

In this study, you'll complete a number of trials which will allow you to choose between two gambles, A or B. You can win real money from the gambles, which will be paid to you at the end of the study. You can't lose money from gambles. Right after you choose each gamble, I'll tell you whether or not you won the gamble.

[For unpaid subjects, the above instead read "In this study, you'll complete a number of trials which will allow you to choose between two gambles, A or B. You can win (imaginary) money from the gambles. You can't lose money from gambles. Right after you choose each gamble, I'll tell you whether or not you won the gamble. At the end of the study, I'll tell you your total winnings. Although no real money will be involved in this study, please try to make your decisions as if the gambles were for real money."]

Here's what the gambles look like:

[An example similar to Figure 1 of the main manuscript.]

The colored bars are just graphical representations of the chance of winning.

Notice that B has a higher chance of paying out. However, B isn't immediately available at the beginning of each trial. It will show as "[Not available yet]". You'll have to wait a random, unpredictable amount of time (sometimes short, sometimes long) for B to become available.

Choosing A will allow you to receive an outcome (either winning or not winning) without waiting, because A is available from the start of each trial. But choosing A won't let you complete the study any faster, because the time you would have waited for B, had you waited for it, will be added to the time you have to wait to get to the next trial (or to the end of the study). Any time you spend waiting before choosing A (although you don't need to wait before choosing A, as you do for B) will be credited towards reducing this wait.

1. Compared to B, A's chance of paying out is
• lower [correct]
• higher
• the same
2. Which gamble gives you more money when you win the gamble?
• A
• B
• They give the same amount of money [correct]
3. Which option can you choose as soon as a trial starts?
• A [correct]
• B
• Either A or B
4. Which option will allow you to complete the study faster?
• A
• B
• Neither; it makes no difference [correct]

Okay, one more thing before we begin.

You'll complete trials in blocks of 3.

[On a blue background:] In odd-numbered blocks (the 1st, 3rd, 5th, and so on), you'll see this background.

[On a yellow background:] In even-numbered blocks (the 2nd, 4th, 6th, and so on), you'll see this background.

The final line of the instructions varied by condition:

• Control: "The task works the same whether you're in an odd block or an even block."
• Within-Pattern: "Within each block, whether even or odd, you can choose either A or B on trial 1, but the task will force you to repeat that choice on trial 2 and trial 3."
• Across-Pattern: "In odd blocks, you can choose either A or B. In even blocks, the task will force you to repeat the series of choices you made in the previous block. So if in the 1st block you chose A, then B, then A, the task will force you in the 2nd block to choose A, then B, then A."
• Force-D: "In odd blocks, you can choose either A or B. In even blocks, the task will force you choose [sic] B on every trial."

References

Ariely, D., & Wertenbroch, K. (2002). Procrastination, deadlines, and performance: Self-control by precommitment. Psychological Science, 13(3), 219–224. doi:10.1111/1467-9280.00441

Barreda-Tarrazona, I., Jaramillo-Gutiérrez, A., Navarro-Martínez, D., & Sabater-Grande, G. (2011). Risk attitude elicitation using a multi-lottery choice task: Real vs. hypothetical incentives. Spanish Journal of Finance and Accounting, 40(152), 613–628. doi:10.1080/02102412.2011.10779713

Bryan, G., Karlan, D., & Nelson, S. (2010). Commitment devices. Annual Review of Economics, 2(1), 671–698. doi:10.1146/annurev.economics.102308.124324

Ciria, L. F., Quintero, M. J., López, F. J., Luque, D., Cobos, P. L., & Morís, J. (2021). Intolerance of uncertainty and decisions about delayed, probabilistic rewards: A replication and extension of Luhmann, C. C., Ishida, K., & Hajcak, G. (2011). PLOS ONE. doi:10.1371/journal.pone.0256210

Giné, X., Karlan, D., & Zinman, J. (2010). Put your money where your butt is: A commitment contract for smoking cessation. American Economic Journal: Applied Economics, 2(4), 213–235. doi:10.1257/app.2.4.213

Hinvest, N. S., & Anderson, I. M. (2010). The effects of real versus hypothetical reward on delay and probability discounting. Quarterly Journal of Experimental Psychology, 63(6), 1072–1084. doi:10.1080/17470210903276350

Horn, S., & Freund, A. M. (2022). Adult age differences in monetary decisions with real and hypothetical reward. Journal of Behavioral Decision Making, 35(2). doi:10.1002/bdm.2253

Luhmann, C. C., Ishida, K., & Hajcak, G. (2011). Intolerance of uncertainty and decisions about delayed, probabilistic rewards. Behavior Therapy, 42, 378–386. doi:10.1016/j.beth.2010.09.002. Retrieved from http://www.psychology.stonybrook.edu/cluhmann-/papers/luhmann-2011-bt.pdf

Rachlin, H. (2014, July 24). If you're so smart, why aren't you happy? Retrieved from http://blog.oup.com/2014/07/psychology-self-monitor-abstract-particular

Rachlin, H. (2016). Self-control based on soft commitment. Behavior Analyst, 39(2), 259–268. doi:10.1007/s40614-016-0054-9

Rachlin, H. (2017). In defense of teleological behaviorism. Journal of Theoretical and Philosophical Psychology, 37(2), 65. doi:10.1037/teo0000060

Rachlin, H. (2019). Group selection in behavioral evolution. Behavioural Processes, 161, 65–72. doi:10.1016/j.beproc.2017.09.005

Read, D., Loewenstein, G., & Kalyanaraman, S. (1999). Mixing virtue and vice: Combining the immediacy effect and the diversification heuristic. Journal of Behavioral Decision Making, 12(4), 257–273. doi:10.1002/(SICI)1099-0771(199912)12:4<257::AID-BDM327>3.0.CO;2-6

Rogers, T., Milkman, K. L., & Volpp, K. G. (2014). Commitment devices: Using initiatives to change behavior. JAMA, 311(20), 2065–2066. doi:10.1001/jama.2014.3485

Siegel, E., & Rachlin, H. (1995). Soft commitment: Self-control achieved by response persistence. Journal of the Experimental Analysis of Behavior, 64(2), 117–128. doi:10.1901/jeab.1995.64-117

Xu, S., Pan, Y., Qu, Z., Fang, Z., Yang, Z., Yang, F., … Rao, H. (2018). Differential effects of real versus hypothetical monetary reward magnitude on risk-taking behavior and brain activity. Scientific Reports, 8. doi:10.1038/s41598-018-21820-0