5 Discussion
The optimality of our actions is constrained by the accuracy of our judgments. Yet, our percepts and evaluations often appear inaccurate. Depending on the context, the exact same alternatives might lead to different choices – seemingly irrelevant contextual information biases our judgments in a consistent and systematic manner. This robust context dependence, thus, poses a puzzle. Why has evolution sculpted the neural system to process information in a manner that reduces the apparent accuracy of our judgments? I argued for the adaptive role of such a mechanism. Adjusting neural responses based on contextual information boosts the efficient encoding and processing of information; contextual information also helps resolve ambiguities in the face of noise and uncertainty.
In this discussion, I will summarize the findings from my doctoral work and highlight their implications for and connections with phenomena outside the scope of perception and economics, and outside the laboratory. My thesis considered three different instances of context-dependent choice across the realms of perception and economics. Each study combined careful experimental investigation of human behavior with simulations using mathematical models of hypothesized computational mechanisms. This approach allowed me to first, chart the parametric influence of context on human choices, and second, arbitrate between different theoretical accounts of the drivers of the observed context dependencies. On the behavioral level, the studies revealed rich patterns of contextual influence spanning interactive, attractive and repulsive effects. On the algorithmic level, my computational work provided evidence that normalization schemes of neural processing can account for many of the observed contextual effects.
5.1 Overview of Results
Chapter 2 introduced decoy effects in economic multialternative choice. It went beyond existing work in the field, which has traditionally focused on three well established decoy effects (attraction, compromise and similarity), and mapped the strength and directionality of decoy influence across the full two-dimensional attribute space. Computational simulations revealed that the rich structure of decoy influence can be produced by a simple, general principle of relative encoding, where attribute values are repulsed away from the context of rival options. In the chapter, I further explored the space of putative normalization schemes and charted their predicted behavioral signatures in the multialternative choice task. Evidence from the choice behavior of a large cohort of human volunteers converged with normalization theories that compressively transduce inputs, appealing to normalization by the central tendency of context and recurrently overweighing the contribution of the target input to the normalization.
Chapter 3 considered the role of this information processing scheme in the domain of perceptual choices along a single visual dimension – orientation. Evidence from three psychophysical experiments indicates that perceptual distractors exert an interactive influence on choices, biasing the accuracy of judgments according to their consistency with target input. This result is in stark contrast with the classic model of distractor influence, which proposes that distractors compete with targets for processing resources and attract choices as a function of the observer’s attention. Instead, I find that focusing attention enhances sensitivity to target stimuli, but does not ameliorate the contextual influence of the distractor. This counter-intuitive finding is consistent with an attention-independent normalization coding scheme, where the gain with which target signals are processed is boosted in consistent contexts, a mechanism which presumably lies outside of voluntary control.
Finally, chapter 4 delved into the computational mechanisms of context-dependent categorization decisions. Through rigorous experimental design, I provide evidence that the observed contextual bias is driven by the combined influence of a normalization-based information processing scheme, continuously adapting to the temporal context of the decision maker, and a Bayesian process, pushing the categorization standards towards the central tendency of context. This finding holds across three distinct stimulus dimensions – lightness, size and numerosity – highlighting the domain-generality of the mechanisms underlying the observed context dependence, and bridging our results with related work on context dependence in evaluations of propositional statements and subjective value.
5.2 Interpretation
Taken together, the results of the three studies in this thesis contribute a unifying account of context dependencies across perceptual and economic decisions. Across all chapters, normalization emerged as a common theme, offering a compelling computational account of the behavioral results. It is perhaps somewhat counter-intuitive that a common neural mechanism would drive context dependencies across all the scenarios I consider here. They do, after all, concern very different stimulus characteristics, spanning visual appearance and subjective value, and pose distinct task demands, including comparing, categorizing and ordering alternatives. Yet, divisive normalization has been identified as a “canonical neural computation,” repeated across brain regions and modalities (Carandini and Heeger 2012). It might, thus, form part of an arsenal of operations that the neural system has evolved to apply across the processing hierarchy for a variety of problems.
The existence of a common computational mechanism underlying relative judgment has long been hypothesized in the psychological literature. Adaptation-level theory (ALT, Helson 1964), for instance, posits that the human brain processes all manner of information in a relative manner. According to ALT, inputs are evaluated against the current adaptation level of the individual, which is shaped by the stimulation present in the temporal and spatial context (residual and background stimuli, respectively, in ALT). ALT strove to provide a unified account of context dependencies across perception, affect, motivation, learning, cognition and interpersonal behavior. Research on contextual effects across these areas has since, however, splintered off into distinct and largely theoretically unrelated research streams, like for instance, work on adaptation and perceptual constancy in perception (Smithson 2005; Webster 2015) and work on the hedonic treadmill in well-being research (Diener, Lucas, and Scollon 2009). The parallels between the underpinnings of context dependencies in perceptual and economic choices highlighted in this thesis signal that there is merit in considering the relativity of human judgment holistically and renewing the search for a domain-general computational mechanism of contextual influence.
5.3 Zooming Out
5.3.1 Context Dependence in the Wild
The instances of context dependence I considered here differ in their definitions of context. Chapters 2 and 3 considered the influence of information present in the spatial context of a decision. In chapter 3, this information was completely irrelevant to the choice problem; the contextual input constituted a distractor stimulus which the decision maker was asked to ignore. In chapter 2, this information formed part of the choice set; the decision maker was asked to evaluate the contextual (decoy) input and assign a preference rank to it. Chapter 4, on the other hand, considered temporal context. The contextual input consisted of stimuli on previous trials. The task demands in the laboratory rendered context “irrelevant” for the evaluation of the imperative stimulus across all three of these definitions. Yet, outside the artificial laboratory setting, where stimuli and trials are independently and randomly sampled, the irrelevance of context is not so clear cut.
Context may provide useful cues for the interpretation of information. The Kuleshov effect, for instance, refers to a phenomenon whereby the same (ambiguous) facial expression can be interpreted as showing different and potentially contradictory emotions depending on the context in which it is encountered. Popularized by filmmaker Lev Kuleshov over a century ago, the effect emerged when the same clip of an actor was stitched together with oppositely valenced scenes, a scene of a funeral or a scene of a child playing. In the former case, the actor’s emotional state was interpreted as melancholy, in the latter – as happiness (Mobbs et al. 2006). It is easy to see why using context as a cue in this situation would be compelling. In daily life, the emotions of the people around us do usually covary with the context in which they are situated. It is primarily in artificial settings, like the psychology laboratory, that facial expressions may be divorced from their surroundings and need to be interpreted following unrelated emotionally charged clips (Mobbs et al. 2006) or photographs (Mullennix, Barber, and Cory 2019).
Context is, thus, useful to make sense of the world. In fact, its role in helping us interpret inputs is so crucial that it features prominently in attempts to construct artificial brains. Context dependencies are inbuilt in various state-of-the-art artificial neural networks. Convolutional networks (CNNs) incorporate information from the spatial context of inputs to analyze and classify images. The base operations of CNNs largely mirror the structure of the visual system, with convolutional and pooling layers serving an analogous computational role as simple and complex cells (Lindsay 2021). Similarly, recurrent networks (RNNs) incorporate information from the temporal context of inputs to interpret information sequences. Analogous processes, where signals are dynamically fed back to processing units, are commonly found in biological brains (Goulas, Damicelli, and Hilgetag 2021). Context dependence is, hence, ubiquitous and crucial across both biological and artificial information processing operations.
5.3.2 Context Dependence as a Tool
Reliance on context is a compelling strategy in the real world, but it can also lead to inconsistent and inaccurate behavioral outcomes. The same processes that I describe in this thesis – decoy (Wu and Cosguner 2020) and repulsive “contrast” effects (Levin 2002) – are often applied by marketing professionals to influence consumer choices. Artificially constructing choice contexts that can tip consumer preferences towards a desired alternative is a lucrative avenue for private companies; recent estimates show that decoy effects can increase sales by as much as 15% (Wu and Cosguner 2020). There is, thus, a clear financial incentive to understand and take advantage of contextual influences on decisions.
Beyond purchasing choices, unintended contextual biases can arise in various social settings, like for instance, the selection of candidates in hiring and admissions decisions (Highhouse 1996; Norton, Vandello, and Darley 2004; Simonsohn and Gino 2013). Raising the stakes does not seem to eliminate the influence of context. Context dependencies sway even the (arguably) most consequential decisions in society concerning policy choices (Herne 1997) and political leadership (Chang, Gershman, and Cikara 2019). Archival electoral data and experimental surveys demonstrate that voting choices in American electoral races systematically violate the normative principle of independence of irrelevant alternatives (Sue O’Curry and Pitts 1995; Chang, Gershman, and Cikara 2019).
Over the past decade, a movement to harness these robust context dependencies “for good” has garnered momentum (Thaler and Sunstein 2008). The central idea of the movement is that policy makers can “nudge” individuals towards better decisions about, for instance, their retirement savings and health care, by designing choice contexts which bias people towards the desired outcomes. This approach is often called libertarian paternalism: manipulating the decision context (e.g. how choice alternatives are ordered or presented) would not affect decision makers’ freedom to choose. It would, however, ensure that if the decision makers were to behave in a context-dependent manner, they would choose the intended option determined by a presumably benevolent entity.
Translating research on choice context dependencies into nudges has been the focus of governmental nudge units, such as the the UK’s Behavioural Insights Team and (previously) the USA’s Social and Behavioral Sciences Team. Economically, nudging approaches are estimated to be several times more cost-effective than traditional public policy interventions, such as tax incentives and other financial inducements (Benartzi et al. 2017). In the domain of retirement savings, for instance, nudging has been implemented to boost enrollment rates, prompt individuals to choose more appropriate retirement plans and increase their savings contributions (Benartzi and Thaler 2007; Iyengar and Kamenica 2010). More recently, there has been an increased interest in applying contextual choice biases to prompt individual action in response to the climate crisis (Schubert 2017; Zaneva and Dumbalska 2020; Carlsson et al. 2021).
The success of green nudges and other similar choice interventions, however, is constrained by our understanding of context dependence in human choice. Despite their ever growing popularity, little is actually known about why and how nudging interventions work (Marteau et al. 2020). Rigorous methodological evaluations, which are largely missing from the field (Marchiori, Adriaanse, and De Ridder 2017; Van Kleef and Trijp 2018), and controlled experiments, like the ones presented in this thesis, are necessary to develop a theoretical understanding of the observed behaviors and mechanisms underlying them. Building that understanding is key for the optimization of choice architecture interventions and the maximization of their impact.
5.4 Conclusion
Why do people do what they do? The behavioral sciences aim to address this big question. The study of decision making is crucial for this endeavor, since our actions are, generally, guided by our decisions. It is particularly important to understand the decision process in those cases when our actions appear inconsistent or inaccurate. Oftentimes, inconsistencies can be traced back to reliance on irrelevant contextual information. Uncovering why context matters for our decisions is important not only as a theoretical venture (e.g. what are the neural mechanisms?, are humans rational?, etc.), but also has important translational implications. Context dependence can be wielded as a tool to protect our best interests, as is the goal of the nudge movement. It may, however, also be applied against those interests. It is, thus, crucial to understand and be aware of manipulations to the choice context.
This thesis contributes to a body of work that aims to further our collective understanding of precisely those processes. My doctoral work compiled evidence on context-dependent choice across the domains of perception and economics. Through careful experimental manipulations, I mapped the behavioral signatures of contextual effects. I combined this data with computer simulations to shed light on the information processing mechanisms underlying context-dependent choice. Taken together, these tools can help uncover the factors and processes driving context dependencies in decision making, bringing us one small step closer to answering the big question.