1 Take it or leave it: Cognitive rules and satisfying choices. Wahida Chowdhury ([email protected]) Institute of Cognitive Science, Carleton University Ottawa, ON, K1S 5B6, Canada Warren Thorngate ([email protected]) Department of Psychology, Carleton University Ottawa, ON, K1S 5B6, Canada Abstract cognitive rules people might use to make choices in these situations. Cognitive rules to make decisions can range from sophisticated evaluations of each alternative, to seeking advice, to flipping a coin. The rules we employ depend not only on our personal characteristics (such as, temperament, cognitive ability, practice, and habit), but also on the decision situations we face (Simon 1956). Psychologists and economists have studied the consequences of varying cognitive rules used in decisionsituations with simultaneous alternatives (e.g., Fasolo, McClelland, & Todd 2007; Gourville, & Soman 2005; Thorngate, 1980). However, the literature for cognitive rules used in TIOLI decision-situations is scarce. Simon (1956) argued that, when choosing an alternative, people tend to satisfice (set a minimum standard of satisfaction and take the first alternative that meets this standard) rather than optimize: (take the best alternative after evaluating all). TIOLI situations force decision makers to use some form of satisficing cognitive rule. We might begin our search by setting a fixed minimum standard: to be acceptable, for example, an alternative must have at least seven out of ten desired features. We might then take the first alternative that meets or exceeds our standard, or we might lower our standard if we examine several alternatives and none meet the standard we set. There are several variations of satisficing rules, prompting a few simple questions: How do variations of the rules compare in the choice outcomes they generate? How simple can cognitive rules be and yet make satisfying choices? We attempt to answer such questions by running a computer model of TIOLI situations, with variations of cognitive rules. We frequently face decision situations (selecting a mate, accepting a job offer, etc.) presenting only one alternative at a time and requiring us to "take it or leave it" (TIOLI). These situations force us to adopt some kind of satisfying rule setting minimum standards and accepting the first alternative that meets or exceeds them. One such rule sets initial standards and does not change them as alternatives are sampled. Another kind of rule modifies the standards in light of sampled alternatives. The present simulation examined how level of initial standards, quality of alternatives, and rule modification speed influenced search length and the quality of choice outcomes. Results show that search length grows exponentially when standards remain fixed and declines drastically even when modifications are slow. Results also support the speculation that people adopting lower standards are far more likely to choose alternatives exceeding their expectations than are people adopting higher standards. Implications for the shift from idealism to realism are discussed. Keywords: decision-making, satisficing rules, search length, take it or leave it Decision-Making Situations The hardest thing to learn in life is which bridge to cross and which to burn. David Russell Most research in human decision-making examines situations where a chooser must select one of two or more alternatives presented simultaneously (e.g., Botti & Hsee 2010; Schwartz 2005). This allows the chooser to compare the alternatives before making a choice. However, many decision-situations present only one alternative at a time, and a chooser is constrained to either take it or leave it (TIOLI). If she/he takes it, no more alternatives are presented; if she/he leaves it, the rejected alternative is not presented again. For example, most of us cannot ask prospective mates to wait for a few years while we search for someone better, or ask a prospective employer to hold a job open for a few months while we look for a better one. Also, we are rarely guaranteed alternatives in the future. We might, for example, reject prospective mates until we are too old to marry, or decline job offers until our skills are no longer required. The present study presents a computer model of TIOLI situations, and investigates the consequences of possible The TIOLI Computer Model Drawing on Thorngate's (2000) original simulation, we wrote a program in "R" that (1) created TIOLI situations and (2) applied variations of satisficing rules for making choices in these situations, and (3) examined how the variations affected the number of alternatives examined and expected outcome. Each TIOLI situation was represented in a matrix of 10 rows and 50,000 columns; its cells filled randomly with 0s and 1s in varying proportions. Each column in the TIOLI matrix 107 2 represented an alternative that could be chosen, and each row represented a feature that was present (=1) or absent (=0). Thus, for example, Column 3 represented the third alternative a chooser would see if she/he rejected the first alternative (column 1) and the second (column 2). If Column 3 contained the vector {0 1 0 1 1 0 0 1 1 0}, it would indicate that alternative 3 had features 2, 4, 5, 8 and 9 (where the 1s were) but not features 1, 3, 6, 7, and 10 (where the 0s were). Setting a maximum of 10 features per alternative was arbitrary; it could have been 5 or 50, but 10 seemed to be a reasonable number. Creating 50,000 alternatives allowed the simulated chooser to reject 49999 of them – a theoretical possibility when standards are high and the chances of alternatives meeting them are low. weight on the number of features in previously encountered alternative). We ran the TIOLI simulation for each of the 6x4x6 = 144 combinations of the three variables (six levels of initial standard * four levels of probability * six levels of weight to current standard). For each combination, our computer programme generated a sample of 0s and 1s in a 10x50,000 TIOLI matrix. Then our simulated chooser examined each of these alternatives, one-by-one, looking for the first alternative that met or exceeded the chooser's initial/adjusted minimum standard. Dependent Variables The programme was iterated 100 times for each of the 144 combinations of independent variables. Each of these 100 iterations contained a new, randomly generated set of 50,000, ten-feature alternatives. When the 100 iterations for each of the 144 combinations of independent variables were completed, the programme printed: 1) the average number of alternatives rejected before finding the first one that met or exceeded the chooser’s current minimum standard, and 2) the average number of features in the accepted alternative. Independent Variables We ran our simulation with many possible combinations of three independent variables. Our first independent variable was the initial setting of a chooser’s minimum standard -- whether, for example, the chooser would take the first alternative with at least 3 features, or at least 4 features, or at least 9 features. We used six levels of this initial setting: 10 (the chooser only accepts an alternative if it has all ten features), 8, 6, 4, 2, and 0 (the chooser has no fixed standard and so accepts the first alternative encountered). Our second independent variable was the probability that a feature would be present – thus varying whether, for example, the expected number of features in any alternative was 5 or 3 or 8. We used four levels of this probability: 0.8 (80% chance that a feature would be present), 0.6, 0.4, and 0.2 (20% chance that a feature would be present). The third independent variable was the rate at which choosers adjusted their minimum standard as they sampled and rejected alternatives. We did this by giving complementary weights to (a) the current minimum standard and (b) the number of features in most recently rejected alternative, where a + b = 1.0. Suppose, for example, a = 0.7 and b = 0.3, and suppose the initial minimum standard of a chooser was 6 out of 10. If the first alternative sampled had only two of 10 features, then the alternative would be rejected and the new minimum standard would be calculated: New standard = (0.7*6) + (0.3*2) = 4.8. If the 2nd alternative had four features, it too would be rejected and a second adjustment of the standard would be made: New standard = (0.7*4.8) + (0.3*4) = 3.48. If the third alternative had five features, it would exceed the new minimum standard of 3.48 and thus be accepted. We used six levels of the ‘a’ weight: 1.0 (100% weight on current standard and 0% weight on the number of features in the most recently rejected alternative), 0.8, 0.6, 0.4, 0.2, and 0 (no weight on current standard, i.e. 100% Cognitive Rules to Make a Choice Different combinations of the third independent variable, rate of adjustment, represented two different cognitive rules of satisficing: When the weights of a = 1.0 and b = 0.0, the chooser never changed her/his initial minimum standard. When the weights of a < 1.0 and b > 0.0, the initial minimum standard was adjusted in light of the number of features encountered as each alternative was sampled and rejected. We report the results separately. Fixed Minimum Standard Average number of alternatives examined We first considered a chooser with a simple, if rigid, cognitive rule: Set an initial minimum standard and stick with it until a choice is made. Figure 1 shows four curves plotting the relationship among minimum standards, probability of features, and search length (average number of alternatives examined). 100000 0.2 10000 0.4 1000 0.6 100 0.8 10 1 0 2 4 6 8 10 Minimum standard Figure 1: Minimum standards, probability of features and search lengths. 108 3 Average number of alternatives examined Visual inspection of Figure 1 shows that search length to find an acceptable alternative increases exponentially with the increase of minimum standards, and that the exponential effect is amplified as the probability of a feature declines. Choosers who stick to high initial minimum standards, and do not adapt to the chances of features in their TIOLI situation, face very long search tasks. For example, a chooser who will accept nothing less than 8 out of 10 features when the probability of a feature is 0.40 can expect to examine about 100 alternatives before finding one with the minimum; an idealistic, uncompromising chooser who wants 10 out of 10 features can expect to examine about 1,000 alternatives before the perfect one comes along. Figure 2 plots the relationship between a chooser’s minimum fixed standard and the average number of features she/he gets in her/his accepted alternative. Prob=1.0 Prob=0.8 Prob=0.4 Prob=0.2 than 5. The chooser would then evaluate the second alternative with the new, minimum standard. Again, if the second alternative met or exceeded the adjusted standard, the chooser would take it; otherwise the chooser would compromise. The chooser would continue this adjustment cycle until she/he finds an alternative that met or exceeded her/his current minimum standard. The rule for adjusting current standard was based on a simple formula: St+1 = ws*St + wf*Ft where St is the current standard employed; St+1 is the adjusted standard employed for the next trial; Ft is the number of features in the most recently rejected alternative; ws is the weight given to the current standard; wf is the weight given to the encountered number of features; and ws+wf = 1.0. Our Fixed Minimum Standard simulation, above, was run by setting ws = 1 and wf = 0, so 100% weight was given to current minimum standard and no weight was given to the number of features in previously encountered alternatives. In this second study we varied ws to be 0.8, 0.6, 0.4, 0.2, and 0.0. Because the chooser adjusts or decreases her/his standard towards the number of features found in encountered alternatives, the lower the weight a chooser gives to her/his current standard, the faster the standard adjusts to new information. Also, the faster the standard adjusts, the fewer the number of alternatives the chooser should search to find an acceptable alternative, and the fewer the number of features the chooser should get in an accepted alternative. We conducted the simulation to examine in more detail the shape of this relationship. Prob=0.6 10 8 6 4 2 0 0 2 4 6 8 Minimum standard 10 Figure 2: Effects of standards and feature probabilities on average number of features obtained. As expected, Figure 2 shows the number of features obtained was equal to or higher than the minimum standard. However, a higher number of features was obtained than the initially set minimum standard as the minimum standard decreased. This suggests that choosers are more likely to get an alternative, with a higher average number of features than what was desired, if their initially set minimum standard was low. Results. Using 0.6 for the probability of a feature, Figure 3 shows six curves plotting the relationship between the initial minimum standard and the average length of search to find an acceptable alternative for various values of ws (weight given to current standard). For comparison, Figure 3 also shows the plot when ws = 1.0, duplicated from Figure 1. Average number of alternatives examined Adjusted Minimum Standard We next considered a chooser with a more complex cognitive rule; the chooser still set a minimum standard before beginning to search, but would adjust the standard in light of the number of features in each rejected alternative. For example, a chooser might begin with a standard that an acceptable alternative must have at least eight out of ten features. If the first encountered alternative meets/exceeds the standard, the chooser takes it without changing the initial standard; if the first alternative has, say, five features, then the chooser would compromise by lowering the standard below 8 but greater 100 1.0 0.8 0.6 0.4 0.2 0.0 10 1 0 2 4 6 Initial standard 109 8 10 4 Figure 3: Effects of initial standards and weights to initial standards on average number of alternatives searched. decline from about 100 to about less than 10 (see Figure 3). This suggests small adjustment to initially set high standard reduces search length considerably but reduces the number of obtained features minimally. Figure 4 also shows that the decline in the average number of features obtained became even less pronounced as the initial standard declined. This suggests again that there is little to be gained by adjusting standards if initially they were set moderate or low. Plots for other probabilities of a feature showed similar shapes and are not reported here for the sake of space. Visual comparison of the top data series (ws =1.0) in Figure 3 with the other data series shows a dramatic decline in the number of alternatives searched when the initially set minimum standard was 10; the values of ws from 0.8 to 0 also reduced search length by an order of magnitude, but not as noticeably as it did when ws decreased from 1.0. This suggests that large reductions in search length do not require rapid adjustments to initial standards or to situational features; a slow adjustment can also decrease search length substantially. Figure 3 also shows that the decline in the search length became less pronounced as the initial standard declined. This suggests there is little to be gained by adjusting standards if initially they were set moderate or low. In order to examine how a decline in ws (weight to current standard) affects the number of features in the chosen alternative, we plotted the average number of features obtained across six levels of the initially set minimum standard (0, 2, 4, 6, 8, and 10) for each of the six levels of ws (1, 0.8, 0.6, 0.4, 0.2, and 0). The plots are shown in Figure 4. As with Figure 3, we report here the plot we obtained when the probability that a feature would be present was 0.6. Average number of features obtained Weight=1.0 Weight=0.6 Weight=0.2 Discussion We created our simulation to explore some consequences of adopting plausible cognitive rules for satisficing (Simon 1956) in take-it-or-leave-it situations. One rule set an initial minimum standard and stuck to it regardless of the chances of desirable features in each alternative examined. A second rule modified the initial minimum standard according to information about the chances of desirable features gleaned from examining rejected alternatives. The simulation results teach at least three lessons about setting standards and adjusting them. The first lesson is that raising standards leads to an exponential increase in the number of alternatives examined before an acceptable one is found. Unless desirable features of alternatives are abundant, perfectionists must be prepared to be very patient. The second lesson suggested by our results is that even slow adjustments of high standards to match more closely the features of alternatives will save considerable search costs. It will also save the disappointment associated with examining additional alternatives that do not meet choosers’ high standards. The third lesson is that lowering standards will not only decrease search costs but also increase the chances of pleasant surprises. Choosers with extremely high standards will face long and costly searches, and will rarely if ever find an alternative that greatly exceeds their standards. Choosers with lower standards will face far shorter and less costly searches, and will frequently choose alternatives that do greatly exceed their standards. Our results suggest that the tradeoffs between search costs and rewards are likely to be optimized by quickly estimating the average number of features in alternatives, then setting a minimum standard somewhere between the average and perfection. The suggestion reflects adjustments in the idealism of youth as life experience nudges them towards realism. It is nicely summarized in the anonymous life prescription, “If all else fails, lower your standards!” Weight=0.8 Weight=0.4 Weight=0.0 12 10 8 6 4 2 0 0 2 4 6 8 Initial Standard 10 Figure 4: Effects of initial standards and weights to initial standards on average number of features obtained. As expected, Figure 4 shows that the average number of features in the chosen alternative declined as ws declined. However, the decline in the number of features obtained was not as dramatic as the decline in search length. For example, Figure 4 shows that, when the initial minimum standard was 10, adjustments of this minimum in light of search experience would reduce the average number of features obtained from 10 to between 6 and 8. However, the average number of alternatives searched would References Botti, S., & Hsee, C. K. (2010). Dazed and confused by choice: How the temporal costs of choice freedom lead to undesirable outcomes. Organizational Behavior and Human Decision Processes, vol. 112 (2), 161–171. 110 5 Fasolo, B., McClelland, G. H., & Todd, P. M. (2007). Escaping the tyranny of choice: when fewer attributes make choice easier. Marketing Theory, vol. 1 (13), 13-26. DOI: 10.1177/1470593107073842. Gourville, J. T., & Soman, D. (2005). Overchoice and Assortment Type: When and Why Variety Backfires. Marketing Science, vol. 24 (3), 382395. Schwartz, B. (2005). The paradox of choice: Why more is less. USA: Harper Perennial. Simon, H. A. (1956). Rational choice and the structure of the environment. Psychological Review; Psychological Review, 63(2), 129. Thorngate, W. (1980). Efficient Decision Heuristics. Behavioral Science, 25(3), 219-225. Thorngate, W. (2000). Teaching social simulation with Matlab. Journal of Artificial Societies and Social Simulation, 3(1). Appendix 1.Download and install from http://www.r-project.org/ the programming language “R” (64-bit). 2.Start R. R Console window will load. 3.Under the “File” tab, select “New script”. An “untitled - R editor” window will open; 4.Copy the entire codes below and paste it into the “untitled - R editor” window. 5.Click the floppy disk icon on top and save the codes under the name “tioli.r”; 6.Close the “R editor” window so that you are back to the “R console” window; 7.Under the “file” tab, select “Source R code” and, then “tioli.r” (the codes you just saved will load in the R Console window); 8.Type in tioli() # InitialStandardNF =How many features did you initially want in an alternative? # prob="what is the probability that features would be present in an alternative? # a=how much weight from 0 to 1 do you give to your standard? # b=how much weight from 0 to 1 do you give to the number of features in current alternative? simulation=function(InitialStandardNF,prob,a){ b=1-a # searchLength is the alternative number that meets your standard # adjStandard is you standard after you adjusted it # CurrentNumFeatures is the number of features in the current alternative. # Set all these variable to 0 before you start your search searchLength=0 adjStandard=0 CurrentNumFeatures=0 # Generate 100 TIOLI situations for (trial in 1:100){ # set your standard as what you specified standardnf= InitialStandardNF # generate the first alternative in a TIOLI situation and see how many features the alternative has. altnum=1 alt=rbinom(10,1,prob) currentnf=sum(alt) # now adjust your set standard depending on how much weight you gave to your standard and find the alternative that meets your standard for(trial in 1:50000){ ifelse(currentnf>=standardnf, (trial=50001), ( (standardnf=a*standardnf+b*currentnf) && (altnum=altnum+1) && (alt=rbinom(10,1,prob)) && (currentnf=sum(alt)) ) ) } #keep a running record of your searchLength for each cycle searchLength=searchLength+altnum #keep a running record of your adjusted standard for each cycle adjStandard=adjStandard+standardnf 111 6 #keep a running record of CurrentNumFeatures for each cycle CurrentNumFeatures=CurrentNumFeatures+currentnf } # report the average across 100 TIOLI situations return(c(InitialStandardNF,prob,a,(searchLength/100),(adjStandard/100),(CurrentNumFeatures/100))) } # run the TIOLI model with the following levels of “InitialStandardNF”, “prob”, and “a” tioli=function() { InitialStandardNF =c(10,8,6,4,2,0) prob=c(0.8,0.6,0.4,0.2) a=c(1.0,0.8,0.6,0.4,0.2,0) print(c("InitialStandardNF","prob","a","searchLength","adjStandard","CurrentNumFeatures")) for(x in InitialStandardNF) { for(y in prob) { for(z in a) { result=simulation(x,y,z) print(c(result)) } } } } #end of "tioli" function 112