Recurrent, robust and scalable patterns underlie human approach and avoidance

Abstract

Background: Approach and avoidance behavior provide a means for assessing the rewarding or aversive value of stimuli, and can be quantified by a keypress procedure whereby subjects work to increase (approach), decrease (avoid), or do nothing about time of exposure to a rewarding/aversive stimulus. To investigate whether approach/avoidance behavior might be governed by quantitative principles that meet engineering criteria for lawfulness and that encode known features of reward/aversion function, we evaluated whether keypress responses toward pictures with potential motivational value produced any regular patterns, such as a trade-off between approach and avoidance, or recurrent lawful patterns as observed with prospect theory.

Methodology/principal findings: Three sets of experiments employed this task with beautiful face images, a standardized set of affective photographs, and pictures of food during controlled states of hunger and satiety. An iterative modeling approach to data identified multiple law-like patterns, based on variables grounded in the individual. These patterns were consistent across stimulus types, robust to noise, describable by a simple power law, and scalable between individuals and groups. Patterns included: (i) a preference trade-off counterbalancing approach and avoidance, (ii) a value function linking preference intensity to uncertainty about preference, and (iii) a saturation function linking preference intensity to its standard deviation, thereby setting limits to both.

Conclusions/significance: These law-like patterns were compatible with critical features of prospect theory, the matching law, and alliesthesia. Furthermore, they appeared consistent with both mean-variance and expected utility approaches to the assessment of risk. Ordering of responses across categories of stimuli demonstrated three properties thought to be relevant for preference-based choice, suggesting these patterns might be grouped together as a relative preference theory. Since variables in these patterns have been associated with reward circuitry structure and function, they may provide a method for quantitative phenotyping of normative and pathological function (e.g., psychiatric illness).

Figure 1. Experimental Design of Keypress Procedure.

(a) This schematic illustrates the four potential responses to the stimuli: to increase, decrease, variably increase or decrease for the same image, or accept the default viewing time of 6 (+2) seconds. The default condition controls for subjects having an intention to keypress and alter viewing time, but not acting on this intention. (b) The traces of individual keypressing behavior to each picture are shown for an anonymous subject. Time intervals are color coded by experimental condition as follows: beautiful female (pink), average female (red), beautiful male (light blue), and average male (dark blue). Stimuli presentation was segregated by gender so that beautiful female faces did not bias all other responses (i.e., responses to male faces). Each blue trace of actual keypress data is shown relative to the default baseline for viewing. (c) Viewing time relative to the default time (location and standard errors) for the ensemble of BF, AF, BM, AM faces are shown as a bar graph.

Figure 2. Preference Uncertainty Trade-off.

(a) shows a graph of (y axis) vs. (x axis) for BF, AF, BM, AM faces in 77 healthy control subjects [experimental conditions (i.e., stimulus categories) are color coded as indicated in (b)]. The central tendency of the manifold is approximated by a black quarter-circle, with its dispersion via crossbars and mathematical formulation as , where N = the number of items in the experimental condition. Spectra for the radial probability distributions of responses to the BF, AF, BM, AM faces are superimposed in (b). Given 20 items for each set of faces, this plot produces a distribution centered on 4.32 bits. In (c), the data for four categories of faces are plotted for one individual.

Figure 3. Interpretation of Trade-off Plot.

This cartoon provides an example of possible keypress patterns that fall at six different positions on the manifold, using data from six subjects toward the same 20 BM faces (F1–F20) for increasing viewing time (data are shown for approach only). For the six approach graphs shown, the x-axis represents the 20 faces in an experimental condition (i.e. BF, AF, BM, AM), and the y-axis represents the number of keypresses toward that face picture. The Shannon entropy was computed using data in this format (see Methods, Analyses, Descriptive Statistical Measures). To schematize the balance of approach and avoidance , one might imagine a matching of graphs #1 with #6, #2 with #5, and #3 with #4, where one graph represents the keypress responses for approach and the other avoidance . For the purposes of illustration, we assigned zero values here to in sub-figure #6. For each sub-figure (#1–6) above and to the right of the manifold, data has been auto-scaled to optimize the pattern display. Overall, this graph represents relative approach or avoidance bias along the polar angle, whereas the extent of indifference/conflict an individual feels toward an experimental condition (i.e. BF, AF, BM, or AM) is distributed in radial fashion from the origin.

Figure 4. Value Function with Group Data.

In (a), the boundary envelope is shown for BF, AF, BM, AM faces in 77 healthy control subjects. The envelope can be fit well either via a logarithmic function or a power function, over the range of keypress responses. As a power function, this envelope has a similar structure to the value function in prospect theory (b). When approach behavior (green points) and avoidance behavior (red points) are plotted together (c), one can readily observe the steeper trajectory of the envelope for avoidance responses, which in prospect theory is interpreted as “loss aversion”. With transformation of the axes (d), both the envelope and envelope show power law scaling.

Figure 5. Value Function with Individual Data.

In (a), data for the BF, AF, BM, AM faces in one individual is shown for and plots, superimposed on the fits for the other individuals in the cohort. With the same log transformation of axes performed for group data, one observes in each individual the signature of a power law. Here, the data for one individual is highlighted (b) above the graphs in lighter colors for the rest of the subjects. It is important to note that the structure of these individual plots is consistent with the respective boundary envelopes for group data. Interpreting rank order of experimental conditions on these graphs depends on how one frames the measurement of relative position, (c). If one frames the ordering of experimental conditions by either axis (dotted blue lines for x-axis, dotted red lines for y-axis), one observes different relative orderings. A third ordering is possible if one frames the positioning relative to the power function fit for (light green line), which calibrates the pattern of responses across items in an experimental condition (H) to the mean intensity of responses (K).

Figure 6. Group and Individual Saturation Plots.

In (a), mean keypress intensity (K) is plotted against standard deviation (σ), for approach and avoidance responses to the BF, AF, BM, AM faces, in 77 control subjects. A quadratic envelope readily fits the avoidance data , and the left side of the approach data distribution for . Most telling are the individual data, where quadratic fits are also observed for each of the 77 individual data sets with the BF, AF, BM, AM data (b). A similar mathematical structure is observed in individual graphs with the BF, AF, BM, AM faces, albeit with different fitting parameters for each of the 77 subjects. These patterns are similar to those reported for ensemble averages of mIPSCs for synaptic GABA_A channels by De Koninck & Mody .

Figure 7. Noise Simulation and Injection for .

(a) Simulation results for variance-matched Gaussian noise (orange dots) do not mimic data from 77 controls over 4 experimental conditions (blue) (also see Figure S9). These simulation data represent alterations in the length of exposure to stimuli, and thus relate to the psychological process of judgment regarding how long to keypress for a stimulus. The minimal overlap between real data and simulated noise is underscored by statistical parametric mapping (i.e., bucket statistics (b)). When the Gaussian noise is injected into the real data, a new manifold is produced (orange dots), which is shifted past the manifold for the Gaussian noise (c). Depending on the noise distributions used for injection into experimental data, one can observe a range of central tendencies for the manifolds resulting from data plus noise, which share features with receiver operating characteristic (ROC) curves (i)–(iv) in (d). The cartoon in (d) can also be compared to Figure 3, where (i) represents the theoretical internal boundary for the trade-off manifold when subjects either keypress to approach or avoid; the central tendency of the experimental data would be (ii), while the outer border with Gaussian noise data would be (iii), and the new manifold due to injected noise would be (iv). The Pflip analysis shown in (e) and (f) allows one to assess the effects of inserting noise into the decision-making process. It specifically alters the valence or polarity of the decision-making shown by experimental subjects for their existing trace profiles in a parametric fashion (i.e., flipping 10%, 20%, 30%, 40%, 50%, etc. of the decisions from approach to avoidance, and vice versa). The graphical effect of this parametric flipping of the valence of decision-making can then be assessed by overlaying graphical representations of existing subject data with representations altered by this decision-making perturbation. In the preference trade-off graph (e), this flipping leads to data convergence toward the midpoint of the theoretical central tendency of the manifold as one goes from 0% flipping to 50% flipping. With 60% to 100% flipping one observes the manifold being stretched back out along its central tendency (i.e., the black line; data not shown). As one goes from 0% to 100% flipping, one effectively reverses the manifold so that it is rotated along the radius line of 45 degrees. In (f), we see that the radial spectra of the Pflip analysis are superimposed and similar across flipping perturbations. The manifold is thus robust to perturbation of the decision-making.

Figure 8. Replication with IAPS Stimuli.

With transformation of the axes in (a) & (b), both the data (red linear fits) and the data (green linear fits) show power law scaling for the individuals in the first and second experiments with the IAPS stimuli. Saturation plots for the same individuals are shown in (c) & (d), where quadratic fits for are shown in red, and for are shown in green.

Figure 9. Individual Data Set from IAPS Experiment.

The nine categories of stimuli used from the IAPS stimuli for these experiments are color coded, and displayed for one example subject. This subject's plot is shown in (a). Their value function, and , is shown in (b), and with log-transformation of K in (c). Note the tightness of the fitted functions in (b) and (c). Tight quadratic fitting is further noted for the saturation function in (d), for both and . Details regarding these fits across the entire cohort of subjects undergoing testing with the IAPS stimuli, can be found in Tables 4 and 5. Note the similar sets of behavioral patterns in this figure to those shown in Figures 2– 6.

Figure 10. Replication with Food Stimuli.

Four types of food stimuli were shown to subjects in hungry and satiated states. The order in which each state occurred was counterbalanced across subjects and separated by approximately one week. These stimuli included pictures of normal colored food, discolored food, prepared food and unprepared food. For the trade-off plot in (a), the center of mass of across these four stimuli differed between the hungry and satiated states, with an angular offset of 58.92° during the hunger condition, and 40.79° during the satiated condition. This difference of 18.13° quantifies the alliesthesia effect, by which homeostatic state can alter the baseline valuation of goal-objects. For these same subjects, the value function for is shown in (b), with the signature of a power law in (c), and approach responses in green and avoidance responses in red. Lastly, the saturation plots for these same subjects are shown in (d), with quadratic fitting of approach responses in green and avoidance responses in red. Color, shape, and open/full coding of the four stimulus types, during hunger or satiation, for approach and avoidance responses, are shown with the same codes in (b)–(d). Details regarding fitting across the cohort of subjects undergoing testing with the food stimuli, can be found in Table 6. Note the similar sets of behavioral patterns in this figure to those shown in Figures 2– 6, 8, and 9.