ANIMAL BEHAVIOUR, 2006, 71, 1119–1129 doi:10.1016/j.anbehav.2005.09.006 What makes information valuable: signal reliability and environmental uncertainty COLLEEN M. MCLINN & DAVID W. STEPHENS Department of Ecology, Evolution and Behavior, University of Minnesota (Received 23 May 2005; initial acceptance 26 July 2005; final acceptance 12 September 2005; published online 10 March 2006; MS. number: A10169) We investigated the roles of signal reliability and environmental uncertainty in animal signal use. We developed a simple model that predicted when animals should switch between choosing the most common option (which we call environment tracking) and following a signal. The model predicts signal following when signal reliability exceeds environmental certainty. We tested this model experimentally using captive blue jays, Cyanocitta cristata. During each trial, the jays had to choose between two stimuli; one option was correct and led to food, and the other was incorrect and provided no reward. In addition, a third stimulus, the signal, provided information about which choice stimulus was correct. Using this procedure, we manipulated signal reliability (the probability that the signal matched the correct stimulus) and environmental uncertainty (the background probability that a given stimulus was correct) in a factorial experiment. Signal reliability and environmental uncertainty influenced signal use roughly as predicted. Jays used the signal when the signal was reliable and the environment was uncertain, and they ignored the signal when it was unreliable and the environment was predictable. Quantitatively, we observed a bias in favour of environment tracking. Jays sometimes ignored the signal when it could have helped them, and they ignored the signal in conditions where signal following and environment tracking produced equal payoffs. We discuss the implications of these findings. Ó 2006 The Association for the Study of Animal Behaviour. Published by Elsevier Ltd. All rights reserved. Many aspects of behaviour, such as foraging, communication and mate choice, involve information processing, and many of the subdisciplines of behavioural science have considered mechanisms of information processing. For example, one can think of most of the basic processes of psychology (e.g. learning, perception, memory) as mechanisms of information processing. Behavioural ecologists have in turn focused on the evolutionary economics of information, asking, for example, whether it pays to attend to or ignore an environmental signal (Stephens 1989; Bradbury & Vehrencamp 2000; Koops 2004). This study considered two factors that should, in theory, influence the value of information: signal reliability and environmental uncertainty. Signal reliability should enhance the value of a signal: the more reliable the information, the more valuable the signal. An example of a reliable signal is the carotenoid-based plumage coloration in male house finches, Carpodacus mexicanus, which accurately indicates nutritional condition during moult, and thus health or foraging ability (Hill & Montgomerie 1994). Females attend to this condition-dependent signal, Correspondence: C. M. McLinn, 100 Ecology Building, 1987 Upper Buford Circle, St Paul, MN 55108, U.S.A. (email: mcli0009@umn.edu). 0003–3472/06/$30.00/0 selecting brighter males and therefore higher-quality mates (Hill 1991). In general, Bradbury & Vehrencamp’s (1998, 2000) models suggest that useful signals should provide a minimum level of accuracy that depends on the relative payoffs of correct and incorrect choices. However, reliability cannot make a signal valuable on its own (Stephens 1989; Bradbury & Vehrencamp 2000). A signal’s value also depends on the statistical distribution of environmental states. Consider a simple situation in which the environment can be in one of two states, A and B, and the best choice for a hypothetical animal is to choose behaviour ‘a’ if the environment is in state A and behaviour ‘b’ if the environment is in state B. When should an animal attend to a signal that tells it whether A or B is true? If A and B are equally likely, then the animal is uncertain about how to behave, and a reliable signal may be valuable. If, however, A is true 99% of the time, and B applies only 1% of the time, then the animal already has very good information about how to behave and a signal may be unimportant, regardless of how reliable it is. We call this environmental uncertainty. Theoretically, then, signal reliability and environmental uncertainty are both required to make a signal valuable. Empirical evidence also supports the potential importance of environmental uncertainty in decision making. 1119 Ó 2006 The Association for the Study of Animal Behaviour. Published by Elsevier Ltd. All rights reserved. ANIMAL BEHAVIOUR, 71, 5 For example, female peacock wrasses, Symphodus tinca, may either search for and spawn with nesting males, or spawn with readily available non-nesting males (Luttbeg & Warner 1999). Females face a trade-off between the cost of searching for nesting males and the increased hatching success of eggs left in nests. Individual female peacock wrasses could increase their fitness by estimating the probability of finding a nesting male and adjusting their reproductive decisions accordingly, and they appear to do so (Luttbeg & Warner 1999). However, peacock wrasses, and animals in general, could make costly mistakes if they rely too heavily on recent experience, especially if the environment changes rapidly. When the environment is uncertain, a signal that reliably indicates the best action could be valuable. Starting from these assumptions, we developed our own model and test of the factors that should make information valuable. always chooses red if p > 1/2 and red is rewarded p of the time. So the payoff function 8 1 > <1 p p < 2 VE ¼ 1 > :p p 2 gives the expected benefits obtained by an environment tracker. This expression gives a simple V-shaped function with a minimum at p ¼ 1/2 (a tracker cannot do better than 50:50 when the environment is a random mix of types), and co-equal maxima at p ¼ 0 and p ¼ 1, where there is complete certainty (Fig. 1). In contrast, a signal follower obtains one unit when the signal correctly indicates the true state, which it does with probability q, and obtains zero when the signal indicates the wrong state, so the payoff function VS ¼ q Model To explore the effects of signal reliability and environmental uncertainty, we developed a simple model using the framework of statistical decision theory (reviewed in Dall et al. 2005). Although we built the model with a specific experimental situation in mind, we think that it captures the general properties of the interaction between reliability and environmental uncertainty. Assume that an animal faces a choice between two alternatives at intervals throughout the day, for example, choosing between a green choice key and a red choice key. On each trial, one key is better than the other, and the animal obtains a food reward if it chooses the good key, and obtains nothing if it chooses the bad key. If, in addition, a third stimulus also provides information about the true state, then we can manipulate the reliability of the signal by controlling how accurately the third stimulus signals the true state, and the environmental uncertainty by controlling the relative frequencies with which red or green are the ‘true state’. Let q represent the signal reliability. A reliable signal matches the true state with probability 1 (q ¼ 1), and an unreliable signal is completely random with respect to the true state (q ¼ 1/2). Let p represent the overall probability that red is the true state. If p ¼ 1/2, there is high environmental uncertainty, because some completely random process determines the true state from one trial to the next. If p ¼ 1 or p ¼ 0, there is no environmental uncertainty (red is always true if p ¼ 1, and green is always true if p ¼ 0). Now consider two simple strategies for dealing with this system. An ‘environment tracker’ always chooses the option that is most commonly ‘true’. So, if p > 1/2, it chooses red, and if p < 1/2, it chooses green. (We assume that the animal has enough experience with the system to know which option is rewarded most frequently). A ‘signal follower’ simply chooses the option that matches the signal. Now we want to calculate the expected gains that individuals playing these two strategies obtain. An environment tracker always chooses green if p < 1/2, and green is rewarded 1 p of the time; similarly, a tracker gives the expected benefits to a signal follower. An animal should prefer signal following to environment tracking when VS > VE. To show how the reliability (q) and uncertainty ( p) parameters determine which strategy is best, we plotted our expressions for VS and VE on the same graph (Fig. 1). These plots show that signal following is best when p is near 1/2 (i.e. when environmental uncertainty is highest), but that the signal-following region increases in size as the signal becomes more reliable. EXPERIMENT The plot of VE is symmetric and our decision to define p as the probability that red is true is arbitrary, so we focused on 1/2 p 1, which captures a range of conditions from complete uncertainty at p ¼ 1/2 to complete certainty at p ¼ 1. Similarly, we restricted the signal reliability VE, payoff to environment tracker 1 Payoff VE or VS 1120 VS, payoff to signal follower Signal following best 0.5 q 0 0.5 1 No uncertainty Highest uncertainty No uncertainty Environmental certainty, p Figure 1. Payoffs for environment tracking (VE) and signal following (VS), plotted as a function of environmental certainty. Signal following is the best strategy near p ¼ 1/2, where VS is greater than VE. MCLINN & STEPHENS: RELIABILITY AND UNCERTAINTY parameter q to the range 1/2 q 1, where q ¼ 1/2 implies a completely unreliable signal and q ¼ 1 implies a completely reliable signal. In this restricted rectangle, signal following is superior to environment tracking whenever q > p. Figure 2 shows the predicted ‘follow signal’ and ‘ignore signal’ regions in the signal reliability– environmental certainty rectangle. Our experiment factorially manipulated signal reliability and environmental certainty using three levels of each. Specifically, we tested environmental certainty values of p ¼ 0.5, p ¼ 0.75 and p ¼ 1.0, and signal reliability values of q ¼ 0.5, q ¼ 0.75 and q ¼ 1.0. These manipulations create nine points on the reliability–certainty rectangle (Fig. 2). Our manipulations created three conditions where we predicted signal following, three where we predicted that subjects should ignore the signal and three where we predicted indifference between signal following and environment tracking. We also tested the same three levels of environmental certainty ( p ¼ 0.5, p ¼ 0.75, p ¼ 1.0) without any signal. These ‘unsignalled’ treatments provide a direct measure of the effect of a signal in signalled treatments. METHODS Colour Assignments Our subjects were six unrelated blue jays, which we had captured as nestlings under appropriate state and federal permits and hand-reared. The birds were of unknown sex and ranged in age from 1 to 3 years at the time of the study. Three of the subjects were na€ıve (band numbers 6, 12, 91), and three had been in other learning experiments or a pilot experiment for this study (band numbers 22, 24, 77). We maintained the subjects in accordance with Least uncertain Experimental treatments 1 Environmental certainty, p Ignore signal track environment On each trial, the birds chose between two coloured pecking keys. Pecking one colour led to food, but pecking the other produced nothing. We used the following pairs of colours: red and green, red and blue, and orange and green. We randomly assigned one colour pair to each subject, and each subject experienced these colours throughout the experiment. To apply our model, we also needed to define the environmental certainty parameter p (the probability that colour X is true). For example, if we assigned the colour pair red and green to a subject, p could be the probability that red is true or the probability that green is true. We determined the meaning of p randomly. Trial Overview 0.75 Follow signal Most uncertain 0.5 0.5 University of Minnesota Institutional Animal Care and Use Committee guidelines, on a 13:11 h light:dark cycle. We conducted the experiment in operant chambers constructed from sheet metal and wood, measuring approximately 61.6 48.3 40.6 cm (Fig. 3). We equipped each box with two panels. The rear panel consisted of a single stimulus light and a perch. We mounted this rear perch on a hinge and connected a microswitch that allowed us to detect when a bird occupied the rear perch. The front panel consisted of three pecking keys (Med Associates ENV 124-AM, St Albans, Vermont, U.S.A.) and a stationary wooden perch. The centre key (slightly elevated above the two side keys) served as the signal or information key. The two side keys served as the response keys. The jays could easily reach the keys from the front perch. A pellet dispenser (Med Associates ENV-203-20) delivered 20-mg pellets into a food cup on the front panel. We connected the entire apparatus to a computer running the Med-PC version IV behavioural test program, Med Associates. A program written in MedState Notation language controlled all of the experimental contingencies and recorded all of the subjects’ responses. 0.75 Signal unreliable 1 Signal completely reliable Signal reliability, q Figure 2. The predicted ‘follow signal’ and ‘ignore signal’ regions on a signal reliability–environmental certainty plot. Signal following is beneficial when q > p. We tested nine experimental treatments (open circles), three each predict signal following, environment tracking and indifference. Although we randomized colour assignments as explained above, we use the example of red and green. A typical trial proceeded as follows. (1) The running program selected the true colour for the next trial, selecting red with probability p and green with probability 1 p. (2) After an 85-s intertrial interval, the computer switched on the rear light, indicating the start of a trial. (3) When the subject hopped on the rear perch, the computercontrolled signal key was then illuminated, showing a white light for 5 s in unsignalled treatments or a red or green light for 5 s in signalled treatments. In signalled treatments, the colour of the signal light matched the true colour (chosen in step 1) with probability q. (4) Next, while the signal light remained illuminated, the computer also switched on the red and green response keys. We randomized the position of the two colours, so that red occurred on the left half of the time and on the right half of the time. (5) The subject’s first peck to one of the response keys indicated its choice and determined the outcome. If the subject pecked the correct key as 1121 ANIMAL BEHAVIOUR, 71, 5 (a) (b) 30.5 cm 40.6 cm 30.5 cm Rear perch Rear stimulus light 48.3 cm Rear perch Water bowl (c) Front perch 48.3 cm 15.2 cm Food cup Signal key Stimulus projectors Pellet dispenser 48.3 cm 40.6 cm 1122 Left response key Right response key Magazine light Food tube Front perch Food cup Figure 3. Diagram of operant box used for the experiment. (a) Overhead view. (b) Rear panel, showing rear stimulus light and hinged ‘start’ perch. (c) Front panel, showing signal key, two response keys, magazine light to signal pellet delivery into cup and fixed front perch. determined in step 1, then the computer extinguished all the lights and delivered food; if it pecked the incorrect key, then the computer extinguished all the lights but delivered nothing. If, after 10 min, the subject had not pecked either key, then the computer aborted the trial and initiated a new one. determine subject behaviour in the absence of a signal, and to look for effects of experience in treatments where one colour is more likely to be rewarded over the course of the experiment. Stability Criterion Treatments The birds experienced three unsignalled treatments and nine signalled treatments. The unsignalled treatments comprised three levels of environmental certainty, p ¼ 0.5, p ¼ 0.75 and p ¼ 1.0 (where p ¼ 0.5 represents complete uncertainty and p ¼ 1.0 means complete certainty, e.g. red is always true). The nine signalled treatments were a factorial combination of the same three levels of environmental certainty with three levels of signal reliability, q ¼ 0.5, q ¼ 0.75 and q ¼ 1.0 (where q ¼ 0.5 represents a signal that is true or false equally often, and q ¼ 1.0 means that the signal is always true). The subjects experienced the treatments in an ABA design: the unsignalled pretest first with the three treatments in random order, followed by the nine signalled treatments in random order, and finally, the unsignalled post-test again, with the three treatments in a new random order. The goal of the unsignalled pretest and post-test was to Within each of the 15 treatments tested, the subjects experienced several hundred trials. Each treatment lasted for a minimum of 300 choice trials and a maximum of 1000. We terminated a treatment when the subject’s behaviour met a stability criterion of less than 10% change in proportional choice of the most common colour over the last three blocks of 50 trials. We examined the data and found no systematic between-subjects variation (from age or experimental history) in the number of trials required to reach the criterion, although there was an effect of treatment. We analysed only data from the final 150 choice trials, when the stability criterion had been met. Forced or No-choice Trials In addition to free-choice trials, 18% of all trials were forced or no-choice trials. In these trials, the computer MCLINN & STEPHENS: RELIABILITY AND UNCERTAINTY illuminated only one choice colour and required a peck to this single illuminated key to end a trial. Forced trials were otherwise like normal trials for the prevailing treatment (i.e. either signalled or unsignalled, and with the appropriate p and q values in effect). In a block of 22 trials, the first four were forced trials (half forced most common trials and half forced least common trials) and the final 18 were free-choice, data-recording trials. This procedure ensured that subjects had recent experience with all possible outcomes before they were allowed to choose freely. P(Correct j Most Common True) and P(Correct j Least Common True) would both be equal to the signal reliability in a given treatment (Fig. 4b). Our graphs and nonparametric analyses used the raw proportions of these two dependent measures, but for parametric tests such as repeated measures analysis of variance, we used arcsine-square-root-transformed proportions to achieve a more normal distribution of the data (Zar 1999). RESULTS Dependent Measures We can view both of the strategies that our experiment examined (environment tracking and signal following) as methods of discriminating good outcomes from bad outcomes. Signal detection theory (Egan 1975; Gescheider 1976) suggests that one can use a truth table to analyse discrimination mechanisms. A simple two-by-two truth table shows that any binary discrimination problem can lead to four possible outcomes (Table 1). In our case, if the ‘most common colour’ is the true state, then the subject can choose the most common colour (correctly) or choose the least common colour (incorrectly). If we know the probability of one of these two events (e.g. r), we can calculate the probability of the other (s) because they are complementary events (i.e. r þ s ¼ 1). Similarly, if the ‘least common colour’ is the true state, then the subject can choose the most common colour (incorrectly) or the least common colour (correctly), and again these are complementary events. We used one probability from each of these pairs as the dependent measures in our study. Specifically, we calculated the relative frequency of correct choices given that the most common colour was true (P(Correct j Most Common True)) and the relative frequency of the correct choice given that the least common colour was true (P(Correct j Least Common True)). Taken together, these two dependent measures give a fairly complete picture of our subjects’ behaviour and separate environment tracking from signal following strategies. Figure 4 shows how plots of P(Correct j Most Common True) and P(Correct j Least Common True) versus signal reliability can distinguish between the environment tracker and signal follower strategies. If the subject tracks the most common colour, P(Correct j Most Common True) would equal 1.0 and P(Correct j Least Common True) would equal 0 across all levels of signal reliability (Fig. 4a). Alternatively, if the subject follows the signal, Table 1. Truth table diagram of dependent measures True state of the environment Subject’s choice Most common Least common Most common Least common P(Correct j Most Common True) P(Incorrect j Most Common True) P(Incorrect j Least Common True) P(Correct j Least Common True) We discuss the results in two parts: predicted and actual behaviour without a signal, and predicted and actual behaviour with a signal. Behaviour without a Signal We tested three unsignalled treatments both before and after the main factorial experiment. Obviously, a jay cannot adopt a ‘signal following’ strategy in this case, because there is no signal to follow. We predicted that an environment tracker should be indifferent when there is environmental uncertainty ( p ¼ 0.5), but it should exclusively choose the most common colour in cases of higher environmental certainty ( p ¼ 0.75 and 1.0). Thus, we should observe P(Correct j Most Common True) ¼ 1.0, and P(Correct j Least Common True) ¼ 0 when p ¼ 0.75 and p ¼ 1.0, but we should observe P(Correct j Most Common True) ¼ P(Correct j Least Common True) ¼ 0.5 when p ¼ 0.5. These predictions require two provisos. First, when p ¼ 1.0, the ‘least common colour’ is never true, so P(Correct j Least Common True) is undefined. Second, when p ¼ 0.5 there is not a ‘most common colour’, although we define one relative to the subject’s most common colour in the other treatments to simplify the analysis. Figure 5 shows an overview of the results with a separate line for each subject. Figure 6 shows the mean values of observed behaviour. Observed behaviour closely matched the environment tracking predictions. P(Correct j Most Common True) was quite close to 1.0 and P(Correct j Least Common True) was quite close to zero for the p ¼ 0.75 and p ¼ 1.0 cases. We did, however, observe a deviation from predictions in our complete uncertainty condition ( p ¼ 0.5). Instead of finding random choice in this condition, as predicted, we found slightly elevated P(Correct j Most Common True) values and reduced P(Correct j Least Common True) values. For the p ¼ 0.5 case, we found a significant difference between the ‘before’ and ‘after’ treatments for the P (Correct j Least Common True) measure (Wilcoxon matched-pairs test: T0.05(2) ¼ 0, N ¼ 6, P ¼ 0.0277), but not for P(Correct j Most Common True) (T0.05(2) ¼ 4, N ¼ 6, P ¼ 0.1730). Taken together, these results show a tendency to choose the ‘most common’ colour even in treatments when it occurred only half of the time. As the before/after comparison shows, this is almost certainly an effect of experience. For example, if the most common colour for a given bird was red, it would have experienced other treatments in which red was the most commonly 1123 ANIMAL BEHAVIOUR, 71, 5 (b) Signal follower (a) Environment tracker 1 P(Correct | Most Common True) Probability 0.5 P(Correct | Most Common True) & P(Correct | Least Common True) P(Correct | Least Common True) 0 0.5 1 0.75 0.5 0.75 Signal reliability, q 1 Signal reliability, q Figure 4. Plots of how P(Correct j Most Common True) and P(Correct j Least Common True) discriminate between environment tracker and signal follower strategies. (a) Predictions for environment tracker: P(Correct j Most Common True) ¼ 1, but P(Correct j Least Common True) ¼ 0. (b) Predictions for signal follower: P(Correct j Most Common True) ¼ P(Correct j Least Common True) ¼ q. rewarded colour, and birds appear to be following this experience in the p ¼ 0.5 case. From an economic perspective, this is a minor violation of our predictions, because in the p ¼ 0.5 treatment there is no reason to prefer one colour over the other, but it costs nothing to show a preference. Overall, the results of our unsignalled treatments show that the jays were superbly sensitive to environmental uncertainty. Our jays capitalized on information that they had gained from experience to efficiently exploit the simple binary choice problem set in our experiment. Behaviour with a Signal To begin, we review the predictions for signalled treatments. An animal following an ‘environment tracker’ (a) Unsignalled pretest P(Correct | Most Common True) or P(Correct | Least Common True) 1124 (b) Unsignalled post-test 1.2 1.2 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 –0.2 0.5 0.75 1 Environmental certainty, p –0.2 Most Common True 6 12 22 24 77 91 Least Common True 6 12 22 24 77 91 0.5 0.75 1 Environmental certainty, p Figure 5. Behaviour without a signal, showing individual variation. P(Correct j Most Common True) and P(Correct j Least Common True) are plotted as solid and dashed lines, respectively, by subject for the three levels of environmental certainty. (a) The unsignalled pretest preceding the main experiment; (b) the unsignalled post-test. P(Correct j Least Common True) is undefined for environmental certainty level p ¼ 1. MCLINN & STEPHENS: RELIABILITY AND UNCERTAINTY P(Correct | Most Common True) or P(Correct | Least Common True) (a) Unsignalled pretest (b) Unsignalled post-test 1.2 1.2 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 –0.2 0.5 0.75 1 Environmental certainty, p –0.2 Most Common True Least Common True 0.5 0.75 1.0 Environmental certainty, p Figure 6. Average behaviour in (a) the unsignalled pretest and (b) post-test, shown with 95% confidence intervals. P(Correct j Most Common True) and P(Correct j Least Common True) are plotted as solid and dashed lines, respectively, representing the mean behaviour for the six subjects. P(Correct j Least Common True) is undefined for environmental certainty level p ¼ 1. strategy always chooses the most common option. So, we predicted that the environment tracker would always choose correctly when the most common colour is true (P(Correct j Most Common True) ¼ 1), and would never choose correctly when the least common colour is true (P(Correct j Least Common True) ¼ 0) (Fig. 4). These predictions apply unambiguously when the relative frequency of the more common option ( p) is strictly greater than 0.5. If p ¼ 0.5, then an animal that chooses to ignore the signal should be indifferent about which option it selects. A signal follower, on the other hand, will be correct whenever the signal correctly indicates the true state. This happens with probability q, so signal followers should choose correctly with probability q. Therefore, for a signal follower, P(Correct j Most Common True) ¼ P(Correct j Least Common True) ¼ q (Fig. 4). Finally, we predicted that signal following is superior to environment tracking when the signal reliability (q) exceeds environmental certainty ( p). Figure 7 gives an overview of the signalled data with a separate line for each subject. Focusing on P(Correct j Most Common True), the figure shows comparatively little variation when the environment is relatively certain ( p ¼ 0.75 and p ¼ 1.0), indicating that most individuals followed an environment-tracking strategy. However, when the environment was completely uncertain ( p ¼ 0.5), we observed much more variation, and a pattern that more closely resembled the behaviour of signal following. We performed two repeated measures ANOVAs to test these patterns statistically. First, we performed a three- by-three factorial repeated measures ANOVA using P(Correct j Most Common True) as the dependent measure. Analysis of the arcsine-square-root-transformed proportions showed a significant interaction between signal reliability and environmental certainty (F4,20 ¼ 8.1793, P ¼ 0.0004). P(Correct j Most Common True) increased with signal reliability when the environment was completely uncertain ( p ¼ 0.5), but changed little with signal reliability in other treatments (Fig. 8). There was also a significant main effect of environmental certainty (F2,10 ¼ 17.109, P ¼ 0.0006), but no main effect of reliability (F2,10 ¼ 3.6776, P ¼ 0.0635). Next, we performed a three-by-two factorial repeated measures ANOVA for the P(Correct j Least Common True) dependent measure (P(Correct j Least Common True) is undefined when p ¼ 1, so we tested three levels of signal reliability and two levels of environmental uncertainty). This analysis showed significant main effects of environmental certainty (F1,5 ¼ 39.9395, P ¼ 0.0015) and signal reliability (F2,10 ¼ 11.9805, P ¼ 0.0022), but no interaction between the treatment factors (F2,10 ¼ 0.5448, P ¼ 0.5962). These main effects are straightforward: P(Correct j Least Common True) increased with signal reliability and decreased with environmental certainty. Figure 9 plots the optimal behaviour and the means of observed behaviour for each treatment. At the extremes of environmental certainty (complete uncertainty, p ¼ 0.5; and complete certainty, p ¼ 1.0), the results agree with our predictions. Jays attended to the signal when the environment was uncertain, and tracked the environment (ignoring the signal) when the environment was certain and 1125 ANIMAL BEHAVIOUR, 71, 5 P(Correct | Most Common True) (a) P(Correct | Least Common True) (b) Environmental certainty p=0.5 1.2 Environmental certainty p=0.75 1.2 Environmental certainty p=1 1.2 1 1 1 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 –0.2 0.75 1 0.5 Signal reliability, q Environmental certainty p=0.5 1.2 –0.2 1 1 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 –0.2 –0.2 0.75 1 0.75 1 Signal reliability, q –0.2 0.5 0.75 1 Signal reliability, q Environmental certainty p=0.75 1.2 0.8 0.5 0.5 6 12 22 24 77 91 0.5 Signal reliability, q 0.75 1 Signal reliability, q Figure 7. Behaviour with a signal, showing individual variation. (a) P(Correct j Most Common True) by subject for each of the nine signalled treatments. Environmental certainty levels are plotted as separate panels, with signal reliability along the X axis. (b) P(Correct j Least Common True) by subject for the signalled treatments. There is no panel for environmental certainty level p ¼ 1, because when the least common colour is never true, the dependent measure P(Correct j Least Common True) is undefined. predictable. We found disagreement with our model in the p ¼ 0.75 treatment, in which the animal’s prior experience provided partial certainty about the true state on any given trial. For this treatment, the model predicts a switch from environment tracking when the signal is unreliable 1.2 P(Correct | Most Common True) 1126 Environmental certainty by signal reliability interaction F4,20=8.1793, P=0.0004 1 0.8 0.6 0.4 0.2 Environmental certainty p=0.5 Environmental certainty p=0.75 Environmental certainty p=1 0 –0.2 0.5 0.75 1 Signal reliability, q Figure 8. Interaction of signal reliability and environmental certainty, for the dependent measure P(Correct j Most Common True). Signal reliability is plotted on the X axis, and the lines represent mean behaviour at a specific level of environmental certainty. Whiskers show 95% confidence intervals. (q ¼ 0.5) to signal following when the signal is completely reliable (q ¼ 1.0), yet the observed pattern resembled environment tracking at all levels of signal reliability. When experience was partially informative, subjects undervalued the signal and overvalued their prior information. DISCUSSION Significance of the Results This study explored the contributions of signal reliability and environmental uncertainty in signal use. Our model predicts that animals should attend to signals when signal reliability exceeds environmental certainty. Qualitatively, our results agree with the model. Signal reliability and environmental uncertainty both influence signal use. The strongest evidence for signal following occurred when the signal was completely reliable (q ¼ 1.0) and the environment was completely uncertain ( p ¼ 0.5). Moreover, as we shifted conditions away from this case, the jays shifted to a strategy of choosing the most frequently rewarded option (environment tracking). We did, however, find a quantitative deviation from our predictions when the signal was completely reliable (q ¼ 1.0) and environmental (or prior) information was partially reliable ( p ¼ 0.75). In this case, the jays’ behaviour more closely resembled environment tracking than signal following: they accepted a 75% success rate even MCLINN & STEPHENS: RELIABILITY AND UNCERTAINTY P(Correct | Most Common True) (a) Environmental certainty p=0.5 1.2 Environmental certainty p=0.75 1.2 Environmental certainty p=1 1.2 1 1 1 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 –0.2 –0.2 –0.2 1 0.5 0.75 Signal reliability, q 1 0.75 0.5 Signal reliability, q 0.5 0.75 1 Signal reliability, q P(Correct | Least Common True) (b) Environmental certainty p=0.5 1.2 Environmental certainty p=0.75 1.2 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 –0.2 –0.2 0.5 0.75 1 Signal reliability, q Optimal Actual 1 0.5 0.75 Signal reliability, q Figure 9. Optimal versus actual behaviour, plotted for each dependent measure, across the nine signalled treatments. (a) Solid lines and closed circles: optimal proportion of P(Correct j Most Common True). Optimal behaviour changes from signal following at environmental certainty level p ¼ 0.5 in the left panel, to environment tracking at p ¼ 1 in the right panel. Dashed lines and open squares: actual P(Correct j Most Common True) values. The lines represent the mean for all six subjects, and the whiskers show 95% confidence intervals. (b) Optimal and actual behaviour for the dependent measure P(Correct j Least Common True) is plotted as above. There is no panel for environmental certainty level p ¼ 1, because P(Correct j Least Common True) is undefined when the most common colour is always true. when they could have achieved a 100% success rate. Overall, our results suggest a bias favouring environment tracking over signal following. For example, we consistently observed environment tracking when environment tracking and signal following provided the same intake rate (i.e. when p ¼ q). This is important because models of signal use and communication typically emphasize the properties of the signals (e.g. Maynard Smith & Harper 1995), but often ignore the simple alternative of exploiting the option that provides the best average payoff. Our results suggest that blue jays are quite sensitive to simple environmental regularities. Our study applied techniques from the psychology laboratory to the economics of information and signal use. Although behavioural ecologists interested in foraging, learning and general decision making routinely use these techniques, students of communication and signalling have tended to take a more naturalistic approach. Although nothing can replace the observation of signals and responses to signals in the wild, we think that our approach offers an important source of supplementary information (reviewed in Rowe & Skelhorn 2004). Using operant techniques, one can directly manipulate economically important variables, and confirm the abilities of animals to respond to them. Results in Context Information-use models have taken several different approaches. One approach focuses on measures of information quantity, such as the Shannon index and other measures of entropy and uncertainty (Shannon & Weaver 1949). Another approach uses the tools of signal detection theory and psychophysics. Signal detection theory comes from engineering (Egan 1975), but psychologists have exploited it in studies of sensation and perception (Gescheider 1976). Our model closely follows a third approach based on statistical decision theory (e.g. DeGroot 1970; Dall et al. 2005), which provides a comprehensive approach to the economics of information use that focuses on how animals will use the information they acquire. Most behavioural models of information use and communication follow the broad outlines of statistical decision theory (e.g. Stephens 1989; Bradbury & Vehrencamp 2000; Koops 2004). Although our model represents a simple example of statistical decision theory (follow a signal if q > p), it addresses two basic concepts: signal reliability and environmental uncertainty. In addition to these theoretical connections, our study also makes connections with several types of empirical work. For example, our study design superficially resembles 1127 1128 ANIMAL BEHAVIOUR, 71, 5 the matching-to-sample paradigm used in studies of animal learning and memory. In matching to sample, the apparatus presents a sample stimulus to the animal (e.g. a red key) and then offers a choice of stimuli in the test phase. The investigators train subjects to select a stimulus in the test phase that matches the sample. The main variable in delayed matching to sample is the delay between presenting the sample and the subject’s choice in the test phase of a trial. Investigators use the effect of delay on performance as a measure of the subject’s memory. Delayed-matching-to-sample studies create an economically simple situation to promote signal use, and use this to study the properties of animal memory (e.g. Blough 1959; Wilkie & Summers 1982; Olson et al. 1995). Our study, in contrast, made minimal demands on memory (because our sample remained illuminated), and asked instead about the economic determinants of signal use. Several studies from a range of disciplines, using a range of approaches, have shown the importance of environmental certainty (our p variable) and signal reliability (our q variable). For example, students of signal detection plot the relation between the frequency of correct acceptances and the frequency of false alarms. They call this relation the receiver operating characteristic (ROC) and consider it to be a fundamental property of the sensory system being studied (Gescheider 1976). Psychophysicists usually manipulate the relative frequency of correct and incorrect stimuli (a p-like variable) to plot the receiver operating characteristic. For example, if correct stimuli are much more common than incorrect stimuli, subjects tend to accept everything, giving a point with relatively frequent false alarms and frequent correct acceptances on the ROC curve (Gescheider 1976). This basic result resembles ours in that animals in certain environments (high p) do not need to use a signal (as in our study) or discriminate carefully (as in the calculation of ROC curves). Behavioural ecologists have also examined the influence of environmental uncertainty on behaviour. Evidence that animals develop representations of variability in the environment that guide their behaviour comes from studies of foraging and habitat use. Devenport & Devenport (1994) showed that least chipmunks, Tamias minimus, and golden-mantled ground squirrels, Spermophilus lateralis, developed temporally weighted estimates of patch quality, allowing them to select patches with the highest probability of reward. Similarly, Cartar (2004) showed that bumblebees (Bombus spp.) used experience with plant quality to guide return foraging trips. Animals may also gain prior probability information by observing the outcome of social interactions, as evidenced by studies of swordtail, Xiphophorus helleri, fighting (Earley & Dugatkin 2002). These studies support our result that animals may track the likelihood of events as one strategy for dealing with environmental uncertainty. Studies from social behaviour have also shown the importance of signal reliability, our q value, in influencing behaviour. Several researchers have shown that animals devalue unreliable signals. For example, Cheney & Seyfarth (1988) used playback experiments to create unreliable individuals in groups of vervet monkeys, Cercopithecus aethiops, and found that group members came to ignore these individuals. Richardson’s ground squirrels, Spermophilus richardsonii, also assess the reliability of individual alarm callers, using it to determine their subsequent time spent in vigilance behaviour (Hare & Atkins 2001). In summary, a large and varied literature addresses communication, information, perception and signalling. We think, however, that our study contributes something by considering how environmental uncertainty and signal reliability combine to influence the economics of information use. Limitations and Future Research Our results seem to indicate that the subjects favoured prior certainty information (derived from experience) over signalled information. We conclude this because subjects used an environment-tracking rule in cases where signal following and environment tracking produced equal payoffs, and because in one case (q ¼ 1.0, p ¼ 0.75) subjects tracked the environment even through they could have done better by following the signal. This observation is important because models of information use tend to underemphasize the importance of simple strategies like environment tracking. It is too early, however, to make strong claims about the generality of our results. Natural signalling systems may or may not have this property, and our procedure may have inadvertently biased our results in this direction. For example, the most common colour (the stimulus colour for which the p ¼ P(Colour X is true) was applied) remained constant for a given subject from one treatment to the next. Thus, for example, across all treatments, a given individual experienced ‘red correct’ more frequently than ‘green correct’, and this made the environment-tracking tactic ‘always choose red’ a good default strategy. In six of the nine treatments of our factorial experiment, this strategy paid off at least as well as any other. The order effect that we found in our unsignalled treatments supports this analysis. One test of this hypothesis would be to arrange a similar experiment where treatments favouring signal following are more common. Our experiment relied on learning. The jays needed to learn the properties of the environment ( p) and the reliability of the signal (q). As our results show, the jays’ behaviour in this system was orderly and broadly consistent with economic principles. However, some critics may think that jays learning about the properties of arbitrary signals in an arbitrary environment have little to tell us about natural information use. Only further study of this problem in more naturalistic settings can resolve this question. We point out, however, that learning is an important component of many natural signalling systems, such as bird song learning (reviewed in Hauser 1996) and alarm-call habituation (Seyfarth & Cheney 1990). More broadly, we will need to study both genetic and experiential factors to have a complete understanding of the mechanisms of animal information use. The results of this study suggest several empirical questions. For example, one might construct a more direct test of the value of information by asking subjects how much they will pay to see a signal. A further step would MCLINN & STEPHENS: RELIABILITY AND UNCERTAINTY involve both a signaller and a receiver in an experiment that combines experimental games with the theory of signalling and information use. We are now pursuing these possibilities. Conclusions Our experiment quantified the effects of signal reliability and environmental uncertainty on animal information use. Our results broadly support an approach to animal information use based on statistical decision theory, showing how blue jays behave in response to manipulation of theoretically important parameters. Our results suggest that prior certainty is fundamental in animal information use, even though students of information use may find it more natural to focus on information reliability. Acknowledgments We thank Matthew Scott and the numerous undergraduate students who helped to complete this research, the behaviour group at the University of Minnesota, Ola Olsson and an anonymous referee. This project was approved by the Institutional Animal Care and Use Committee at the University of Minnesota (Animal Subjects Code 0301A40421). Funding for C.M.M. was provided by a Dayton and Wilkie Fellowship from the Bell Museum of Natural History, and by the Department of Ecology, Evolution, and Behavior, University of Minnesota. References Blough, D. S. 1959. Delayed matching in the pigeon. Journal of the Experimental Analysis of Behavior, 2, 151–160. Bradbury, J. W. & Vehrencamp, S. L. 1998. Principles of Animal Communication. Sunderland, Massachusetts: Sinauer. Bradbury, J. W. & Vehrencamp, S. L. 2000. Economic models of animal communication. Animal Behaviour, 59, 259–268. Cartar, R. V. 2004. Resource tracking by bumble bees: responses to plant-level differences in quality. Ecology, 85, 2764–2771. Cheney, D. L. & Seyfarth, R. M. 1988. Assessment of meaning and the detection of unreliable signals by vervet monkeys. Animal Behaviour, 36, 477–486. Dall, S. R. X., Giraldeau, L.-A., Olsson, O., McNamara, J. M. & Stephens, D. W. 2005. Information and its use by animals in evolutionary ecology. Trends in Ecology and Evolution, 20, 187–193. DeGroot, M. H. 1970. Optimal Statistical Decisions. New York: McGraw-Hill. Devenport, L. D. & Devenport, J. A. 1994. Time-dependent averaging of foraging information in least chipmunks and goldenmantled ground squirrels. Animal Behaviour, 47, 787–802. Earley, R. L. & Dugatkin, L. A. 2002. Eavesdropping on visual cues in green swordtail (Xiphophorus helleri) fights: a case for networking. Proceedings of the Royal Society of London, Series B, 269, 943–952. Egan, J. P. 1975. Signal Detection Theory and ROC Analysis. New York: Academic Press. Gescheider, G. A. 1976. Psychophysics: Method and Theory. Hillsdale, New Jersey: L. Erlbaum. Hare, J. F. & Atkins, B. A. 2001. The squirrel that cried wolf: reliability detection by juvenile Richardson’s ground squirrels (Spermophilus richardsonii). Behavioral Ecology and Sociobiology, 51, 108–112. Hauser, M. D. 1996. The Evolution of Communication. Cambridge, Massachusetts: MIT Press. Hill, G. E. 1991. Plumage coloration is a sexually selected indicator of male quality. Nature, 350, 337–339. Hill, G. E. & Montgomerie, R. 1994. Plumage colour signals nutritional condition in the house finch. Proceedings of the Royal Society of London, Series B, 258, 47–52. Koops, M. A. 2004. Reliability and the value of information. Animal Behaviour, 67, 103–111. Luttbeg, B. & Warner, R. R. 1999. Reproductive decision-making by female peacock wrasses: flexible versus fixed behavioral rules in variable environments. Behavioral Ecology, 10, 666–674. Maynard Smith, J. & Harper, D. G. C. 1995. Animal signals: models and terminology. Journal of Theoretical Biology, 177, 305–311. Olson, D. J., Kamil, A. C., Balda, R. P. & Nims, P. J. 1995. Performance of four seed-caching corvid species in operant tests of nonspatial and spatial memory. Journal of Comparative Psychology, 109, 173–181. Rowe, C. & Skelhorn, J. 2004. Avian psychology and communication. Proceedings of the Royal Society of London, Series B, 271, 1435–1442. Seyfarth, R. & Cheney, D. 1990. The assessment by vervet monkeys of their own and another species’ alarm calls. Animal Behaviour, 40, 754–764. Shannon, C. E. & Weaver, W. 1949. The Mathematical Theory of Communication. Urbana: University of Illinois Press. Stephens, D. W. 1989. Variance and the value of information. American Naturalist, 134, 128–140. Wilkie, D. M. & Summers, R. J. 1982. Pigeons’ spatial memory: factors affecting delayed matching of key location. Journal of the Experimental Analysis of Behavior, 37, 45–56. Zar, J. H. 1999. Biostatistical Analysis. Upper Saddle River, New Jersey: Prentice Hall. 1129