An understanding of the nature of psychological, and neuroscientific, explanations may help clarify the difference between psychology and neuroscience. As an editor, I have repeatedly had to deal with papers that seemed to me to provide explanations that should not count as psychological, or even scientific. These cases forced me to think about what would count. I have come to think that psychological explanations follow certain templates (schemata), basic forms with details to be filled in.
The following templates are tools of psychology that are used for prediction, control, and understanding of behavior and mental states. (I discuss some neuroscience templates later.) For this purpose, they must refer to states that are observable (for prediction) or manipulable (for control). The following templates account for most or all of the explanations I find acceptable (at least for cognitive psychology), but I have no way to be sure that I have not missed something big. These templates do not vie with each other in the way that competing theories sometimes do; they may coexist and continue to be useful for different purposes. Historically, however, many of them have arisen in response to difficulties that other templates had in producing convincing accounts of certain phenomena.
In listing these templates, I want to make three points. First, psychology is a limited field, defined by the types of explanations it provides. Other fields of inquiry, such as economics and sociology, are also concerned with human behavior and experiences, but they have different tools for explanation (with some overlap). Second, these templates work. They succeed in providing explanations that are useful for prediction, control, or understanding. Third, they are scientific. Explanations in psychology may be tested, criticized, improved, and even rejected. In sum, psychology is a science, like many others. It does not need to be replaced by neuroscience, just as chemistry does not need to be replaced by physics.
The oldest template may be the idea of association, which was apparent in the writings of philosophers as early as Aristotle. Simple associations may be seen as a network of connections among psychological objects of some sort, with each connection characterized fully by two objects and a strength. Usually the objects were mental representations. Carl Jung, following many 19th century psychologists, used a word-association test as a way to probe the contents of minds, and the idea of association was central to Freud’s theory of dreams, slips of the tongue, and hysterical symptoms.
The basic idea is that two objects are associated as a result of experience, perceived similarity, or some innate connection. When one object is later stimulated in some way, it evokes the other. Stimulation may result from the presentation of a stimulus, and the effect may be observed through its effect on some response.
A modification of the basic idea was to assume that there are different kinds of associations. For words, these might include super-ordinates, sub-ordinates, co-ordinates, complements, synonyms, opposites, and so on. People could limit reported associations to one type or the other. The idea of labeled connections between nodes potentially increases the capacity for representation, so long as the use of a network of associations was not limited to spreading activation from node to node, determined only by the strength of their connections. Thus, the idea of association could provide some basis for other templates to be described.
The simpler idea of association has been an essential component of explanations of memory since the beginning. It is still very much alive in psychology, particularly in such effects as semantic priming (Schvaneveldt & Meyer, 1973) and in priming studies that fill the journals. For example, the call to prayer is associated with prosocial behavior (Duhaime, 2015). Often such studies rely on what might be considered very weak associations involving chains rather than direct connections. The results are often difficult to replicate. However, the basic idea of semantic priming has been repeatedly confirmed.
Stimulus–organism–response: additive models
For Pavlov and others, the associations of interest were not (just) between mental representations but also those between stimuli and responses. Pavlov inspired more sophisticated theories of learning based on the idea, such as that of Hull (1943). The Rescorla–Wagner (1972) theory of learning builds on these earlier theories and goes beyond them, still retaining stimulus–response and stimulus–stimulus associations as primary tools of explanation. Much of this work and theorizing was on the conditions under which such associations were formed. It was assumed that many of them (Pavlov’s reflexes) were innate.
Along the way, people started to think of simple stimulus–response associations, with nothing in between, as insufficient. The classic textbook by Woodworth and Schlossberg (1954; I read the 1960 revision for my introductory psychology class) notes that the organism also needs to be considered, so its standard model was S-O-R. (They used “organism” rather than “subject” because they were already using S for stimuli!) The basic framework here is still that of association, but with two steps: that between stimuli and some internal representation of it, which is usually also affected by characteristics (temporary or long-lasting) of the organism, and that between the internal representation and its expression in a response. Moreover, the internal representations can combine inputs from two or more stimuli into a single representation.
Brunswik used such an approach in studies of perception, which led to the early work of Kenneth Hammond (Dhami & Mumpower, 2018). The contemporary study of judgment—in which a subject must produce a quantitative evaluative response when presented with a multi-attribute stimulus—is full of explanations of this type. The common idea was that cues (features or attributes) of stimuli in the world were first represented internally, then weighed and added up to produce a perception or judgment.Footnote 3
One slightly extended application of this approach is found in the work of Michael Birnbaum and others who study quantitative judgments in which the cues are also quantitative (e.g., Birnbaum & Veit, 1974). The general idea is that in a judgment task, two monotonic transformations occur. One of these is between a stimulus and its internal representation. Internal representations are then added with weights, or subtracted (for judgments of differences), to produce a summary representation, which is then transformed by another function into a response, mapping a feeling of intensity to (e.g.) a number. The response-producing function can differ for different responses, even if they come from the same internal representations, thus producing scale convergence.
A related approach is conjoint measurement, in which internal representations are inferred from comparative responses (e.g., Krantz, Luce, Suppes & Tversky, 1971; Tversky, 1967). In sum, a standard form of psychological explanation accounts for the connection between stimuli and responses in terms of weighted additive combinations of (internal representations of) cues in the stimuli.
In part as a reaction to the Pavlovian tradition, some psychologists (e.g., Irwin, 1971; Skinner, 1938) started to emphasize a different type of explanation, often attributed to Thorndike (1911), who proposed the law of effect to account for the effect of reward on the frequency of behavior leading to the reward (analogous to the role of reproductive success in the theory of evolution). We can explain behavior by its consequences. Resistance to this idea came from the feeling that this was not a casual explanation, since causes of behavior must precede the behavior. However, this was not really a problem because the general approach obviously required some sort of prior learning or some other reason for an animal to represent the connection between behavior and its consequences.
This approach found full expression in the work of the (self-described) behaviorist B. F. Skinner, but the general approach is not limited to behaviorist explanation. It can be extended to cognitive explanations, in which choices are explained in terms of beliefs and desires. Beliefs can be affected by learning but can arise in other ways. Desires or goals include those resulting from drives, but these too can arise in other ways. The general idea is that we explain S’s choice of C by showing that “S chooses C because S believes that C will more likely lead to outcome O, which S prefers to alternative outcomes.” The slots that need to be filled in are those concerning beliefs and desires/goals.
This sort of explanation is common in the study of decision-making, which also refers to a normative standard, a tool for evaluating decisions as better or worse, in which beliefs are represented as probabilities of particular outcomes, and desires for outcomes are represented as their utilities (Baron, 2008). This kind of normative model is used in other fields, for example, cost-effectiveness analysis in medicine. It is sometimes used as a descriptive model in economics.
Starting perhaps with Wiener’s Cybernetics (1948), an approach developed that explained behavior in terms of the flow of information. As computers became more powerful in the 1950s and 1960s, this approach became common in cognitive psychology, with several variants. One variant involved modeling thought with computer programs. Other variants involved flow diagrams, which could be seen as extensions of the S-O-R approach described above, with more boxes in the middle between S and R, and with the boxes doing more computing than could easily be understood in terms of networks of associations, addition, and subtraction. Still other versions of this approach involved mathematical models of the uptake and use of information.
Information processing encountered at least two sources of resistance. One was to the idea that a model was an explanation. There is something to be said for this objection. An explanation should increase understanding, but a model can (in principle) be accurate yet beyond understanding itself. Most models are understandable. That said, I occasionally see papers where the model is mathematical, a bunch of equations, and, while the model “works,” examination of the equations yields no understanding of why it works, and therefore no clear predictions about where else it would work aside from the data presented. It might as well be a black box with a person inside who, in fact, acts like a person, hence “fits the data” from what other people do in the same situations.
The second early objection came from behaviorism. Behaviorists objected to the postulation of internal representations because they were not directly observable. Although it might appear that observations of the brain may provide evidence of those internal states, it is difficult (as I argue later) to identify brain events with the postulated internal states in an information-processing model. Ericsson and Simon (1980) argued (convincingly to most cognitive psychologists) that, with some care, verbal reports could serve as evidence of the postulated states.
In information-processing models, the slots that need to be filled in are essentially the same ones needed for a description of a computer algorithm, that is, a flow chart. Each step requires inputs in the form of some representations, and an output, which may serve as an input to another step. In addition, the step must be described in terms of a function that produces outputs from the inputs. Importantly, some steps may require conditional decisions: “If this, do that.” Such flow charts (sometimes described only verbally) have been used extensively in the study of heuristics for decision-making and many other areas.
What makes a good template?
Successful templates work because they correspond to reality, at least in cognitive psychology. We cannot observe associations directly, just as we cannot observe electrons.Footnote 4 However, in some sense, we know that they exist. We can create them and then show in many ways that they have been created. Similarly, people (and probably many other animals) often make decisions by considering the possible outcomes and the relative value that they have, as well as their certainty of occurring. Moreover, the flow of information in tasks such as reading can be broken down and observed piece by piece. These templates are not in themselves theories or hypotheses, but each one provides a language and set of assumptions, within which theories may be stated and questions may be asked. For example: Are associations symmetric? Does information flow in one direction only or is there feedback from higher (later) levels to lower ones? Which functions operate in parallel and which in series? The corresponding questions can be answered in ways that allow prediction, control, and understanding.Footnote 5
A second property of a good template is that it is adaptable to applied psychology and the questions that it poses. If we want to influence people’s decisions, the simplest way (not always possible) is to change their values for outcomes or their beliefs about which outcomes are likely to occur as a result of each option. In education, we can create useful associations and avoid creating harmful ones. If we know how information is processed, we can discover where errors are generated and thus, find ways to reduce them. Conditioned responses (of Pavlov’s sort) have been part of behavior therapy since the outset.
An example is the field called judgment and decision-making, which is based on a design consisting of three models (Baron 2012): normative, prescriptive, and descriptive. Normative models specify standard by which judgments and decisions are evaluated. Sometimes they are just “distance from the right answer,” for example, in the case of numerical judgments of quantity, but they may also involve probability theory or expected utility theory. Descriptive models attempt to explain how people do the tasks in question, with special attention to the relation between their responses and the normative model. Prescriptive models are plans for applications. Based on the results of normative and descriptive models, they are designed to promote judgments that are closer to the normative model. The descriptive models are within psychology, but clearly the concepts that they employ must make some reference to the terms of the normative models, such as subjective probability or utility, and they must also refer to objects that can function in prescriptive models.
The concepts of subjective probability and utility are imposed on the situation, much like longitude and clock time. The earth did not come with north–south stripes on it, and the mind (or brain, for that matter) does not come with a probability meter. Yet we can study how probability judgments are related to relative frequencies, whether they are consistent with each other, and how probabilities can be best communicated. We have, in fact, learned quite a bit about these issues. Likewise, in the study of decision-making under risk, for example, extensive analysis has been done of the various ways people make use of information about value and probability information in combination.
Some templates are banging the door to be let in to the list of standard ones (as others have done in the past). One that I think should not be admitted involves the use of metaphor. A comparison of metaphor with more useful templates might help to elucidate their advantages.
An example of the use of metaphor is construal level theory (Trope & Liberman, 2010), as it is applied by many researchers.Footnote 6 The idea is that thinking processes can be characterized as high or low in their level. The high/low dimension corresponds metaphorically to several others: abstract/concrete, distant/nearby (in time, space, or social distance), and holistic/detailed. It is often applied to situations where the interpretation is less obvious, for example, by claiming that one attribute of a stimulus is higher than another (when the attributes could be things like price and quality, or dollars and probability). A typical experiment might involve manipulating or measuring one of these attributes and looking for a relation with another. For example, an increase in perceived distance might lead to more holistic processing.
It may be possible to make sense of several parts of this account. For example, representations that are distant are likely to be less detailed. However, such an account would not explain other inferences from the theory, such as the effect of level (high/low) on the discount rate, or on holistic vs. analytic. Which of these is higher? Analytic thinking may be more detailed (lower), but it may also be more cognitively advanced (hence, higher). Abstract thinking may be higher than concrete thinking, but we sometimes speak of deep thought, which would be lower.
Moreover, it is unclear why manipulation of perceived distance should affect the type of thinking that is done. The associations of dimensions are not assumed to operate through the usual association mechanisms (discussed earlier). It is not as if the word “future” evokes the concept “abstract,” which then, in turn, causes the subject to think abstractly. The idea that “the dimensions are associated one way because most people see them that way” does not seem to me to answer the question of why temporal distance increases abstract thought, if it does. Metaphors may lead to predictions, but it is hard to see how they lead to understanding.