Meta-Psychology, 2023, vol 7, MP.2021.2911 https://doi.org/10.15626/MP.2021.2911 Article type: Original Article Published under the CC-BY4.0 license Open data: Not Applicable Open materials: Not Applicable Open and reproducible analysis: Not Applicable Open reviews and editorial process: Yes Preregistration: No Edited by: Rickard Carlsson Reviewed by: Peder Isager, Julia Rohrer Analysis reproduced by: Not Applicable All supplementary files can be accessed at OSF: https://doi.org/10.17605/OSF.IO/CAFU5 Visual Argument Structure Tool (VAST) Version 1.0 Daniel Leising, Oliver Grenke, and Marcos Cramer Technische Universität Dresden We present the first version of the Visual Argument Structure Tool (VAST), which may be used for jointly visualizing the semantic, conceptual, empirical and reasoning rela- tionships that constitute arguments. Its primary purpose is to promote exactness and comprehensiveness in systematic thinking. The system distinguishes between concepts and the words (“names”) that may be used to refer to them. It also distinguishes various ways in which concepts may be related to one another (causation, conceptual implication, prediction, transformation, reasoning), and all of these from beliefs as to whether something IS the case and/or OUGHT to be the case. Using these elements, the system allows for formalizations of narrative argument components at any level of vagueness vs. precision that is deemed possible and/or necessary. This latter feature may make the system particularly useful for attaining greater theoretical specificity in the humanities, and for bridging the gap between the humanities and the “harder” sciences. However, VAST may also be used outside of science, to capture argument structures in e.g., legal analyses, media reports, belief systems, and debates. Keywords: Modelling, Formalization, Narrative, Theory, Humanities, Science Introduction Argument structures are ubiquitous and important: We encounter them every day, in newspaper articles, in court rulings, in political and non-political debates on and off screen, and in our own more informal con- versations with other people, and with ourselves. In an abstract sense, most arguments are about what is true, why and how things are related to one another, and about whether things are good or not. Too often, however, arguments seem to run in circles, end in stale- mates, or just fizzle out and are given up upon, instead of being conclusively resolved. We argue (yes), that these things happen because people tend to lose sight of some of the claims that they or others have made before. Therefore, it is often advisable to aim for a comprehen- sive analysis of all relevant argument components. An- other reason why so many arguments remain unproduc- tive is that those who argue tend to overlook the actual complexity of their own and others’ claims, and the rel- ative vagueness of many claims. Interestingly, the same problems seem to plague much of the more “narrative” theorizing that is so com- mon in the humanities, and in psychology. In fact, there has been no shortage of calls for better (e.g. more for- malized) theorizing in psychology, precisely to counter these shortcomings (Devezer et al., 2021; Eronen & Bringmann, 2021; Fried, 2020; Glöckner & Betsch, 2011; Muthukrishna & Henrich, 2019; Robinaugh et al., 2021; Smaldino, 2017, 2019). Concrete advice on how exactly such better theorizing may be achieved is largely missing, however (Borsboom et al., 2021). In the present paper, we introduce a tool devised for dealing with those problems. The tool is called VAST (Visual Argument Structure Tool). The core idea is to visually display all relevant components of an argument structure at once, while at the same time aiming for ex- actness. A comprehensive display will make it harder to overlook or downplay related claims made earlier (e.g., because those previous claims do not align well with more recent ones). Visual displays may also be more intuitive and easier to digest for most users, especially when compared to the alternative of using algebraic ex- pressions. After all, there is a reason why so many ar- ticles in scientific journals as well as in the news me- dia are accompanied by figures illustrating their main points. Furthermore, visual displays tend to be more parsimonious: With formulae, the same variable name will have to be written again each time it is used as an input or output of some new equation. In contrast, a visual display may incorporate the variable only once, and establish all of the relevant relationships with other variables through arrows or lines pointing in various di- rections. This is how the matter is handled in VAST, and aligns well with the typical approaches in Structural Equation Modelling (SEM), Structural Causal Models (SCM) and Directed Acyclic Graphs (DAG) (Dablander, 2020; Pearl, 1995; Pearl & Mackenzie, 2018; Rohrer, 2018). The graphical structuring of arguments was also inspired to some extent by developments in formal ar- https://doi.org/10.15626/MP.2021.2911 https://doi.org/10.17605/OSF.IO/CAFU5 2 Visual Argument Structure Tool (VAST) Version 1.0 1 Table 1. Components of the System Element Name Meaning Symbolized by Default Range Concept A feature that may apply to certain objects Frame with abstract label 0; 1 Name A natural-language label that may be used for a concept Frame with label in quotation marks 0; 1 Higher-Order Concept A specific combination of elements that may - in its entirety - apply to certain objects Frame containing two or more elements 0; 1 Data A set of actual observations Frame with thick black edge on one side 0; 1 IS How much X is the case Pentagon containing IS in capitals 0; 1 OUGHT How much X ought to be the case Pentagon containing OUGHT in capitals 0; 1 Perspective How much perspective-holder agrees with X Oval connected to IS / OUGHT Containing name of perspective-holder 0; 1 Naming How appropriate it is to call X by the name Y Arrow accompanied by lowercase letter n -1; 1 Conceptual Implication How much thinking of something as being X also implies thinking of it as being Y Arrow accompanied by lowercase letter i -1; 1 Causation How reliably X will trigger Y Arrow accompanied by lowercase letter c -1; 1 Transformation How strongly X maps onto Y Arrow accompanied by lowercase letter t -1; 1 Prediction How well Y may be predicted from X Arrow accompanied by lowercase letter p -1; 1 Reasoning How much X is a reason to believe Y Arrow accompanied by lowercase letter r -1; 1 gumentation (see Baroni et al., 2018). VAST overlaps significantly with all of these previ- ous approaches, and also incorporates many elements of formal logic, in particular logical connectives (AND, OR and XOR) from classical propositional logic (Bün- ing & Lettmann, 1999) in the tradition of Boole (1854) and Frege (1879). Given that we allow truth-values between 0 and 1 (see below), VAST is also influenced by continuously-valued logics (Preparata & Yeh, 1972). However, VAST is comparatively broader and more inte- grative in that it explicitly accounts for various types of relationships between concepts (i.e., naming, concep- tual implication, causation, prediction, transformation, and reasoning). The strength of all of these relation- ships may be expressed in terms of the same metric, as we will explain in more detail below. VAST also ac- counts for the possibility that concepts may be applied to different sets of objects, for claims as to whether something IS and/or OUGHT to be the case, and for dif- ferent perspectives on these issues. We discuss some of the overlap and differences between VAST and previous tools with a similar scope further below. The System In this section, we introduce the different types of elements that, taken together, constitute our system in its entirety. Table 1 lists all of these elements alongside each other. To facilitate comprehension, we will use a variety of examples along the way to illustrate their po- tential uses. Concepts Concepts are the basic building blocks of cognition. Note that concepts are assumed to exist before language is being used (see below) — they may exert their influ- ence irrespective of the words (“names”, see below) that are used to denote them. A concept assigns values to ob- jects. Thus, concepts are very similar to mathematical functions in that they produce an output value for every input (i.e., object). In the most simple case, that output will be dichotomous, so the concept will yield a value of either 0 or 1 for each object. Here, 1 means that the object is an exemplar of the concept, and 0 means that the object is not an exemplar of the concept. For instance, a person may look at a number of objects and determine whether any of them are exemplars of the concept that one would refer to with the word “car”. 3 Note, however, that many concepts are not dichotomous but allow for continuous variation of output values. In VAST, this variation is usually normalized to a range be- tween 0 and 1, to make comparisons between different concepts easier. This basically incorporates the so-called “prototype approach” to classification (Rosch, 1978), in which an object may be a more or less typical exemplar of a category / concept / class. In VAST, concepts are usually displayed in the form of frames bearing abstract labels (e.g., X, Y, or XYZ). It is important to note that the display of a concept only sym- bolises the assumed existence and relevance of a cogni- tive process that would assign certain values to some objects and other values to other objects. It is agnostic in regard to the desired, assumed, or measured distribu- tion of those values. The question of whether something is or should be the case is the realm of IS and OUGHT statements, which will be introduced later. Empirical data constitute a special type of concept, which is symbolized by adding a thick black edge to one side of the respective frame. This is basically the same distinction that is made between “manifest” (measured) and “latent” (imagined) variables in Structural Equation Modelling (see Figure 12 and the accompanying text for an example). Types of Relationships Between Concepts The next major element of the system is the relation- ship that may exist between concepts. These are basi- cally IF-THEN relations: IF one concept does apply (to some of the relevant objects), THEN another concept also applies, at least to some extent. There are differ- ent qualities of such relationships, however, which must be distinguished from another. In VAST, we explicitly account for naming, implication, causation, prediction, transformation, and reasoning relationships, which we consider to be among the most common and relevant ones. Specifying additional relationship types is also possible, if needed. All of this will now be explained in more detail. Relationship Type 1: Naming (n). People tend to use words to denote the different ways in which they think about objects. In VAST, these words are called “names”. To clearly distinguish concepts and their names from one another, VAST uses abstract la- bels (e.g., C, R) for the former and real language labels in quotation marks (e.g., “Cat”, “Rocket”) for the latter. The relationship between a concept and a word denot- ing that concept is called a “naming relationship”. It is symbolized by an arrow pointing from the concept to the name, accompanied by the lower letter n. As with all relationship types (see above), the arrow stands for XY „Bat“ „Furious“„Enraged“ Z „Device used for hitting a ball in some games“ „Small, nocturnal flying animal with leathery wings“ n n n n n n Figure 1 Selected naming relationships. The word "bat" is a homonym for concepts X and Y, whereas the words "fu- rious" and "enraged" are synonyms for concept Z an IF-THEN relation: IF an object is an exemplar of the respective concept, THEN one may call this object by the respective name. Figure 1 displays some examples of naming relation- ships, including homonyms (the same name is used for different concepts) and synonyms (different names are used for the same concept). Note that both (a) the ap- propriateness and (b) the strengths of the relationships displayed in Figure 1 are treated as irrelevant for now. Distinguishing between concepts and their names is often necessary, because idiosyncratic word usage ac- counts for all sorts of problems (e.g., misunderstand- ings) in everyday arguments. The same issue is relevant for psychology, which continues to suffer from — often unacknowledged — jingle-fallacies (use of homonymic theoretical concepts) and jangle-fallacies (use of syn- onymic theoretical concepts) (Block, 1995). Relationship Type 2: Conceptual Implication (i). Conceptual implication is about the extent to which classifying objects as exemplars of one concept implies also categorizing the same objects as exemplars of an- other concept. Figure 2 displays an example. Here, when an object is considered to be a “sun”, the same object is also somewhat likely to be considered “hot” and “bright”. Conceptual implications are symbolized by arrows accompanied by the lowercase letter i. Figure 2 contains three different ways of displaying basically the same information. In the display on the left-hand side, all concepts and the relevant relation- ships between them are displayed as a part of one co- herent whole. This is the default mode that we suggest for use with most VAST analyses, as it maximises parsi- mony while retaining all of the relevant information. In the middle display, the conceptual implications (type i) 4 S „Sun“ H B „Hot“ „Bright“ n i i n n Sun Hot Bright i i S H B i i S „Sun“ n B „Bright“ n H „Hot“ n FIMM Figure 2 Conceptual implications among three concepts. Left: full (default) mode in which concepts and names are connected within the same structure. Middle: full (default) mode with concepts separated from their names. Right: Finger-is- Moon-Mode (FIMM) in which concept labels reference the concepts’ names. The dashed lines symbolise that these are alternative ways of displaying the same set of concepts, names, and relationships, not three parts of the same VAST display among the three concepts, and the naming relationships (n) have been separated from one another. This way of displaying things may sometimes be helpful to avoid clutter. However, this approach comes at the price of somewhat lower parsimony because every concept now has to be displayed twice. In the display on the right- hand side, we use concept labels that directly reference the concepts’ names. This is what we call the “Finger-is- Moon-Mode” (FIMM), as it abolishes the explicit distinc- tion between signifier and referent. This constitutes an- other possible way of reducing clutter, but comes at the significant risk of overlooking the importance of seman- tics, especially (partial) homonymity, synonomity, and antonymity. For example, another display using FIMM could show that the concept Star has the same con- ceptual implications (Hot, Bright) as the concept Sun. Here, the use of FIMM might obscure the fact that this is the case simply because “Sun” and “Star” are two differ- ent words for the exact same type of thing (S). To high- light the risk of semantic ambiguities like this one, we recommend explicating when the FIMM is being used, by adding the respective acronym in one corner of the display (see Figure 2). Also, many concept names are too long to be used as concept labels. In these cases, we recommend the approach exemplified in the middle of Figure 2. As a next step, we will introduce four more types of relationships between concepts that frequently fea- ture in argument structures. Figure 3 displays the ways in which they are distinguished from one another (in terms of lowercase letters accompanying the respective arrows), along with a very simple example for each type. Note that, for simplicity, this figure uses FIMM, as signalled by the acronym in the upper right-hand corner. Relationship Type 3: Causation (c). Many important articles and books have been writ- ten about causation (Eronen & Bringmann, 2021; Pearl, 1995; Pearl & Mackenzie, 2018; Rohrer, 2018). In VAST, we use a concept of causation that is also re- flected in how most experimentalists tend to think about their research designs. This concept involves temporal order as a necessary ingredient: Causes always precede effects, but never the other way round. Also, causation would become evident if we were able to manipulate the suspected cause variable and then observe subse- 5 Smoking Lung Cancer Temperature in Celsius Temperature in Fahrenheit c t Height Number of Y Chromosomes p X + 4 = 8 X = 4 r FIMM Figure 3 Four more types of relationships between concepts (c = causation, t = transformation, p = prediction, r = rea- soning) quent changes in another variable. Note that all of this concerns the ways in which we (and most people, pre- sumably) think about causation, irrespective of whether such a suspected causal link may ever be proven or dis- proved in terms of data. Note also that most causal rela- tionships between concepts could be — but do not have to be — decomposed into a number of intervening steps. For example, the causal relationship in Figure 3 reflects a relatively proximal link between cause (Smoking) and effect (Lung Cancer). It could be amended by inserting Tar Accumulation as a mediator that is caused by Smok- ing and that causes Lung Cancer. Relationship Type 4: Transformation (t). This relationship type is used to account for situa- tions in which the applicability of one concept may be deduced from another concept by mere computation. The respective example in Figure 3 reflects a case in which one variable (Temperature in Celsius) is basi- cally rescaled into another variable, by multiplying the former’s values with a factor (1.8) and then adding a constant (32). The specific values for the factor and the constant are not displayed, but could be displayed. The transformation type of relationship may also be used to account for scoring procedures, such as the spe- cific ways in which an operational measure of socio- economic status is derived from a number of indicators (e.g., highest degree attained, annual income). Relationship Type 5: Prediction (p). This type of relationship is about knowing something about the values of Y when we know something about the values of X. Note that this is possible without know- ing anything about the mechanism underlying the asso- ciation. For example, type p relationships may ignore the direction of causal effects, as in the respective ex- ample in Figure 3: Here, a person’s Height predicts the Number of Y Chromosomes that same person has, al- though certainly the former is not the cause for the lat- ter. Often, such predictive relationships may be found and described first (e.g., a certain set of symptoms ap- pearing together in patients), and only later be replaced by more specific explanations (e.g., in terms of a virus causing all of those symptoms). Relationship Type 6: Reasoning (r). This relationship type is about the conclusions that people draw from certain premises, on purely intellec- tual grounds. It reflects the idea that if some concept applies (e.g., X + 4 = 8, see Figure 3), one may infer that some other concept (e.g., X = 4, see Figure 3) also applies. Note that this is not limited to conclusions that would generally be regarded “logical”, but to just about any conclusion that someone thinks they may draw. In fact, VAST may be used to first explicate one person’s line of reasoning and then refute that reasoning based on some other reasoning. For example, Peter may think that the results of some empirical study clearly suggest that Y is the case, whereas Trudy may think otherwise. Such discrepancies may then be explained using VAST, by analysing why exactly Peter and Trudy come to such different conclusions (e.g., because one of them trusts the authors of the study, whereas the other does not). So-called “logical conclusions” simply constitute a spe- cial case in which certain lines of reasoning are viewed as (in-)defensible by a group of people (e.g., scientists) who endorse some set of reasoning rules. That endorse- ment then serves as the premise for drawing conclusions as to whether “X is reason to believe Y” or not — which is another type r relationship. Additional Relationship Types. In the present paper, we only address those types of relationships between concepts that we think feature prominently in many arguments — everyday ones as well as scientific ones. Needless to say, the selection is and has to be somewhat subjective. It is relatively easy to come up with examples of other relationship types that may be useful to employ under certain cir- cumstances (e.g., metamorphosis (m): when X turns into Y over time; association (a): when thinking of X makes it likely to also think of Y; element of (e): when X is among the ingredients that, together, constitute Y etc.). We assume that the principles laid out in the following (e.g., regarding relationship strength and the 6 X Y X Y X 1) 7) 2) 4) 8) c „weak“ c 0.5 X Y3) * Y = 2 sin(X) + 750 X YOR Z X Y c 0.3 c c * X c 0.3 c X Y c 0.3 9) c <> 0 c 0 X Y c 0.3 10) c ? X5) Y* Z c * Y = 2.45 X + Z Y X6) Y AND Z c1 c3 c2 Figure 4 Relationship strengths and relationship patterns construction of higher-order concepts) will still apply in these instances. In cases where a relationship between concepts is assumed to exist but the exact nature of that relationship is (yet) unknown, we recommend using the letter u. Relationship Strength In VAST, the default interpretation of an arrow that points from one concept (e.g., X) to another (e.g., Y) is that this relationship is considered relevant and positive (i.e., the more X the more Y). Thus, if an arrow is absent between X and Y, this means that the relationship is zero and/or that it is regarded unimportant for the present analysis. So far, we did not use any further specifica- tions of relationship strength, and this approach may be perfectly sufficient in many cases. Sometimes, however, such specifications will be deemed useful or even nec- essary. VAST allows for the use of verbal labels such as “weak”, “strong”, “negative” etc. for this purpose. This approach will often be appropriate when trying to visu- alise the structure of an existing argument that has been made using the natural language. It may also be the most useful approach when a numerical specification seems not possible (yet). Case 1 in Figure 4 displays an example. Note that we rather arbitrarily added the letter c (causation) to all the arrows in Figure 4. This may easily be replaced with any other relationship type, as everything we say here about relationship strength applies equally to all types. If a simple numerical quantification of relationship strength is wanted, we propose using normalized co- efficients ranging from −1 to 1. “Normalized” means that these coefficients ignore the particular scales of the concepts that they connect, but rather quantify relation- ship strength in terms of proportions of these features’ ranges. Note that this is only possible if such a range may reasonably be assumed to exist. Based on our own experience, this seems to be the case with most psycho- 7 Min(X) Max(X) Min(Y) Max(Y) Min(X) Max(X) Min(Y) Max(Y) Min(X) Max(X) Min(Y) Max(Y) Min(X) Max(X) Min(Y) Max(Y) Min(X) Max(X) Min(Y) Max(Y) Coeff = 1.0 Coeff = 1.0 Coeff = 0.5 Coeff = 0.5 Coeff = 0.5 Figure 5 Relationship strength quantified in terms of default coefficients ranging from -1 to 1. Upper two examples: relationships between two dichotomous concepts. Lower three examples: relationships between two ordinal or metric concepts logical concepts, however. The default coefficient of relationship strength that we propose reflects the increase in Y (expressed as a percentage of the available range of that feature) that is associated with a “perfect” or “complete” increase in X (i.e, an increase that covers the whole available range of X, from the smallest conceivable value to the high- est conceivable value). Case 2 in Figure 4 presents an example in which the coefficient is 0.5. This type of coefficient may be applied to relation- ships between dichotomous concepts as well as to re- lationships between continuous concepts. In the for- mer (dichotomous) case, it may be interpreted as the percentage of cases in which a complete change in X (from 0 to 1) is accompanied by a change in Y (from 0 to 1). In the latter (continuous) case, the coefficient reflects the average increase on the continuous Y scale that accompanies the largest possible increase (from 0 to 1) on the continuous X scale (again both expressed in terms of percentages of the respective ranges). Figure 5 contains a number of examples showcasing this broad applicability. Negative coefficients are to be interpreted accordingly: the more X is the case, the less Y is the case. This default coefficient of relationship strength is generic enough to be applied to all types of relation- ships between concepts (e.g., Type p: “wearing glasses” makes it 70 percent likely for a person to also be “smart”; Type r: It is 90 percent reasonable to assume someone “is in love with you” when that person “giggles a lot while talking to you”; Type c: being “obese” makes it 50 percent likely for someone to develop “Diabetes Type II” as a consequence). For non-dichotomous concepts (see the lower three panels of Figure 5), the default coefficient of relation- ship strength is largely agnostic regarding the distribu- tions of concepts’ values: For example, the relationships displayed in the middle panel and in the panel to the right have the same strength coefficient (0.5) but in the latter case the relationship is deterministic whereas in the former case it is noisier. This difference may also be accounted for in VAST, as we will discuss in the next section ("Noise"). VAST’s default coefficient of relationship strength is based on percentages of the ranges of the concepts that the relationship connects. Sometimes, however, there may be good reasons to deviate from the default (e.g., when earthquake magnitude on the unbounded Richter 8 scale is part of an argument). In such cases, using other measures of relationship strength (e.g., an exact func- tion translating X into Y) is possible. As the exact func- tion connecting certain concepts will often be too long to be written above an arrow in its entirety, we suggest placing it somewhere else in the display and referencing it using an asterisk (e.g., Case 3 in Figure 4). Diamonds should be used when several concepts are jointly related to another concept. This includes logical connectives such as AND, OR and XOR (exclusive OR), as shown in Case 4 in Figure 4. When a more specific formula is needed to derive a joint output from several inputs (e.g., a scoring procedure), the diamond and as- terisk elements may be combined, as shown in Case 5. A diamond with AND inside it may also be used to sym- bolise the interaction effect that two concepts (X and Y) have on a third concept (Z). Case 6 in Figure 4 displays this possibility along with the two main effects of X and Y on Z, so this is basically a VAST-type depiction of two- factor ANOVA. We also use Case 6 to showcase the pos- sibility of indexing coefficients (c1, c2, c3). Doing so is often useful to facilitate discussions among analysts. Noise Sometimes, we may not only wish to display the relationship between specific concepts, but acknowl- edge that there are additional unspecified influences (“noise”) on these concepts, as well. To symbolise these influences, we recommend using “noise arrows” similar to the ones that are used in Structural Equation Mod- elling (SEM). A noise arrow always points toward a given concept (i.e., the one that is affected by the noise) but does not originate in a specific concept. Cases 7, 8, 9 and 10 in Figure 4 provide examples. Note that, other than in Structural Equation Modelling, noise ar- rows in VAST do not stand for the residuals that remain between observed Y values and the Y values that one predicts from X. Rather, they stand for other influences apart from X that may move the values of Y toward its maximum (default, positive coefficient), or toward its minimum (negative coefficient). Case 7 displays the de- fault situation in which noise may lead to an increase in concept Y. If noise may move the values of a concept in either direction, this may be specified using “<> 0”, as shown in Case 8. If it is important to specify that there is no noise, this may be expressed using a coefficient of zero, as shown in Case 9. This expresses the idea that the only factor whose values may make a differ- ence with regard to the values of concept Y is concept X. Finally, when an influence may exist and be relevant for the analysis, but one is not really sure yet, we recom- mend using a question mark as “coefficient” (see Case 10 in Figure 4). TIC „Temperature in Celsius“ n t1 t2 TIF „Temperature in Fahrenheit“ n BH„Body height“ n p1 p2 BW „Body weight“ n B„Bird“ n i1 i2 A „Animal“ n X„Huge“ n Y „Gigantic“ n i 1 GW„Global warming“ n c1 c2 GM „Glaciers melting“ n PNA „Peter has no alibi“ n r2 PG „Peter is guilty“ n r1 Figure 6 Relationship arrows connecting concepts both ways. Use of two arrows implies that the exact ways in which the concepts are related depend on direction. Use of a bidi- rectional arrow (as in the last case) implies that direction does not matter. The i 1 coefficient in the latter case re- flects the assumption that concepts X and Y are identical Relationship Direction Arrows in VAST stand for IF-THEN relationships be- tween concepts. We will now briefly address the direc- tion into which arrows may point, and how this differs between relationship types. Generally speaking, if there is an arrow pointing from X to Y, there may also be an arrow pointing from Y to X. In most cases, the shape of the respective relationship will differ depending on its direction. If both directions are of interest to the cur- rent analysis, we recommend signalling this difference by using two separate arrows, one for each direction. Figure 6 presents a few examples. A few specifics need to be briefly discussed in this regard: First, naming relationships are special in that arrows may only point from a concept to its name but not the other way round. Second, when the strengths of conceptual implications (type i) between X and Y dif- fer depending on direction, this means that one con- cept (the one that is the target of the arrow with the higher coefficient) is broader and more inclusive than the other. This is highly relevant to all kinds of concep- 9 „Expected Increase of average global temperature in the next 50 years, measured in Kelvin“ IS OUGHT 2 0 INC n Figure 7 A simple example with IS- and OUGHT-statements regard- ing the same concept tual hierarchies (taxonomies). Third, VAST does allow for causal (type c) arrows pointing from X to Y and back. This is relevant for displaying all kinds of positive and negative feedback loops. Note that this diverges from recommendations in the literature on Directed Acyclic Graphs (Rohrer, 2018). Fourth, VAST also allows for displays of circular reasoning (type r), because the pur- pose of VAST is not to prescribe rules as to how one should think, but to make visible the ways in which someone thinks. This includes the possibility of display- ing beliefs that others may find unconvincing or even irrational. Fifth, we recommend using a bidirectional arrow with an “i 1” coefficient for expressing the idea that X and Y are identical (see Figure 6). Likewise, a bidirectional arrow with a coefficient of “i -1” would im- ply that one concept is the exact opposite of the other (e.g., between concept H named “huge” and concept T named “tiny”). The IS and OUGHT elements In many arguments, the extent to which something is considered to be the case and the extent to which something should be the case play important roles. To capture these extents, VAST uses two special elements, called IS and OUGHT. Both denote specific values on a given concept. They are important in a variety of ways: First, disagreements often arise because people start from different premises regarding the extent to which something IS the case (e.g., whether vaccines are safe) or the extent to which something OUGHT to be the case (e.g., whether one should trust the government). Second, discrepancies between IS and OUGHT-values on the same concept often explain why people decide to act in certain ways — often they do so in order to move the IS-value closer to the OUGHT-value. In VAST, IS and OUGHT are symbolized by pentagons which include the respective term (IS or OUGHT) in capitals. These are connected to one or more concepts using simple lines rather than arrows, in order to distin- guish them from relationships between concepts. The specific IS- and OUGHT-values are written next to the respective lines. Figure 7 shows a very simple example. IS and OUGHT are not concepts themselves but rather denote specific locations within the range of val- ues that a given concept may have. If IS/OUGHT is used, specifying the metric of measurement may some- times be helpful (e.g., Figure 7). If it is not possible or useful to specify this metric, we recommend express- ing IS and OUGHT values in terms of fractions of the normalized range (between 0 and 1) of the respective concept. If no specific value is given, we recommend using “applies more likely than not" (> 0.5) as the de- fault interpretation. Note that IS may be interpreted as a measure of the respective concept’s central tendency (e.g., the arith- metic mean). It is possible, however, to provide whole ranges of IS- or OUGHT-values, if a single value is deemed insufficient. X Y „Having a good day“ „Being in a good mood“ c Concept Name Unit IS OUGHT M SD M SD X „Having a good day“ % 50 10 70 10 Y „Being in a good mood“ % 65 25 65 10 Figure 8 An example of how assumed and desired concept features may be displayed separately in a table, to avoid clutter Going further, it may sometimes be helpful to specify the assumed and/or desired distribution characteristics of a concept even more. In such cases, we recommend providing the respective information in a separate Table on the side, to avoid clutter. An example is shown in Figure 8. The example tells us that, if a person (= ob- ject) “has a good day”, that person will be more likely to also “be in a good mood”, and that this effect is a causal one. The table also tells us that (a) it would be good if the average percentage of people having a good day would increase (from IS: 50 to OUGHT: 70), and that (b) the variation (SD) in this percentage would go down (from IS: 25 to OUGHT: 10). The latter goal may 10 HPW IS OUGHT Daniel Marcos 0 4 Analysts: Daniel, Marcos 0.9 0.8 0.8IS OUGHT 01.0 0 „Hours per week that Daniel spends studying Esperanto“ n Figure 9 Different IS and OUGHT values for different perspective- holders be rooted e.g., in the assumption that it is important to avoid extreme unhappiness in people. The Perspective Element Almost by definition, arguments tend to involve dis- agreements between viewpoints. To account for this, VAST incorporates a so-called “perspective” element that reflects how strongly a given entity agrees with something. This “entity” is usually a person, but it may also be a group of people or something more abstract like a corporation. The perspective element may only be used to con- dition IS and OUGHT statements. When used to condition an IS-statement, it reflects the extent to which a perspective-holder agrees that the given level of the concept applies. When used to condition an OUGHT-statement, it reflects the extent to which the perspective-holder agrees that this is the most desirable level of the concept. Levels of agreement may be quanti- fied using values between 0 ("does not agree at all") and 1 ("agrees completely"). If a perspective-holder’s level of agreement is left unspecified, we recommend using "tends to agree" (> 0.5) as the default interpretation. If a (part of a) VAST display does not explicate who the perspective-holder is, then IS and OUGHT state- ments reflect the view of the analyst who created the display (see next section). Figure 9 showcases the use of the perspective el- ement in VAST: The name of the perspective-holder is displayed inside an oval, and the strength of the perspective-holder’s belief is again expressed using co- efficients ranging from 0 to 1. Note that a value of 0 would only imply a complete lack of agreement with X, not the belief that the opposite of X is true. This is necessary because many of the concepts that we use in everyday life do not have clearly defined opposites. If it seems necessary to not only visualise a perspective- holder’s lack of agreement with X but also what else (Y) they believe in, that alternative view will have to be specified, as well. The specific example given in Figure 9 conveys a wealth of information at one glance: Daniel and Marcos both think that Daniel does not spend any time studying Esperanto. However, Daniel is perfectly certain about that (1.0) whereas Marcos — who cannot really know for sure — is a little less certain (0.8). Also, Daniel is almost certain (0.9) that he should not take up any Esperanto-learning, whereas Marcos is also quite sure that Daniel should spend 4 hours per week learning Es- peranto. Making differences such as these visible may go a long way in explaining the different behavioural choices that people make. The perspective element of VAST incorporates the im- portant issue of subjective certainty, which plays a key role in scientific theorizing. In fact, if one plotted all of the possible IS-values (X-axis) against a perspective- holder’s subjective certainties (Y-axis) and rescaled the latter such that their sum is 1, one would basically ob- tain a density distribution very much akin to a Bayes prior. Outside of scientific theorizing, however, inspect- ing entire distributions of possible IS-values is relatively rare. Thus, we recommend displaying the IS-value with the highest subjective certainty as a default. If neces- sary, alternative IS-values and their respective certain- ties may be displayed in addition. The perspective fea- ture may also be used to express someone’s “hunches” (e.g., the suspicion that there may be another yet unrec- ognized factor involved in accounting for some effect). This also includes suspicions as to what someone may or may not have meant by saying something (i.e., impli- cations). The Analyst Element Each VAST display has to be created by someone. Notably, the persons creating such displays are respon- sible for arranging the various concepts and their re- lationships with one another in the most accurate or helpful ways possible, but not for the actual content of the respective argument. In fact, it is possible to display the structure of an argument with great pre- cision while at the same time disagreeing wholeheart- edly with most or all of the points that are being made. This is why we prefer to call these persons “analysts” rather than “authors”. We recommend naming the an- alyst who created a display in a header, as also shown in the upper left corner of Figure 9. The persons named as analysts are responsible for the display in its entirety (again: irrespective of how much they agree with the display’s actual content). To make this point clear, we decided to include the same persons (Daniel and Mar- 11 X Y1 c ISIS Tina Toby 1.0 0.0 0.5 0.5Z c X Y2 c Z c Analyst: Peter Figure 10 Higher-order concepts. Here, IS and OUGHT statements for different perspective-holders refer to two higher-order concepts, each of which contains a number of causal rela- tionships between lower-order concepts cos) in Figure 9, in two different roles: as analysts, and as perspective-holders. Altogether, the Figure tells us that Daniel and Marcos (analysts) agree that Daniel and Marcos (perspective-holders) have different views on how much time Daniel should spend learning Esperanto (the specifics of their viewpoints were discussed above and will not be repeated here). In other words, they “agree to disagree”. Higher Order Concepts After having introduced all of the different ways in which concepts may be related to one another, as well as the IS, OUGHT, perspective and analyst elements, it is now time to introduce a final element of great im- portance: In VAST, any combination of elements may itself become a “higher-order concept” (HOC) and thus be related to other (higher-order) concepts or be the subject of IS or OUGHT statements. In this, all of the rules explained so far do apply as well. Figure 10 displays a very simple example. Here, two persons (Toby and Tina) are portrayed (by Peter) as dis- agreeing in regard to the question of whether the causal effect of X on Z is mediated by Y1, or by Y2. Toby is certain that the former is the case, whereas Tina is un- decided. Similar displays may be used to account for all sorts of differences between viewpoints, such as naming conventions (i.e., what it IS that a given label refers to, or what that label OUGHT to be used for). Such issues are as greatly relevant to contemporary psychology as they were decades ago (Block, 1995). It is important to note that a higher-order concept binds all of its components together in an inseparable fashion: The higher-order concept applies (to some ob- ject) if and only if all of its lower-order components ap- ply. By using higher-order concepts, VAST users may “zoom in” on certain parts of a display if they wish to X B Y c c „Antecedent“ „Consequence“„Black Box“ n n n B M1 M2 i 1 „Mediator 2“„Mediator 1“ c IS 1 n n Analyst: Imani Figure 11 Using higher-order concepts to “zoom in” on a particular part of a VAST display. The lower part of the display de- tails what concept B is about elaborate on its details, or “zoom out” when they de- cide to rather ignore some of the details for some time. Figure 11 provides an example: The lower part of the figure “zooms in” on the meaning of one of the concepts (B) that features in the causal chain displayed in the upper part of the figure. Specifically, we learn that B stands for M1 being the case (IS) and M1 eliciting M2. Furthermore, all of this elicits Y. Higher-order concepts may be used to display a hi- erarchy of concepts that apply to different sets of ob- jects. So far, we abstained from explicating the sets of objects that the concepts in a VAST analysis apply to. Instead, we tacitly assumed that those objects were the same across all concepts. Sometimes, however, specify- ing the sets of objects to which different concepts apply is necessary. We recommend using Greek letters for this purpose. Figure 12 displays a hypothetical example. Here, in- telligence test scores and school grades were obtained from students (τ). Remember that the thick black edges of the respective concept frames symbolise the fact that these concepts were actually measured. We assume that the students’ intelligence test scores do predict 12 SEL(σ) INT(τ) GRA(τ) INT(τ) GRA(τ) p1 0.5 p2 0.2 „Students‘ Intelligence-Test Results“ „Students‘ Grades“ HOC1(σ) HOC2(σ) „Selectiveness of School“ INT(τ) GRA(τ) SEL(σ) cc n1 n2 n3 HOC1(σ) HOC2(σ) „Effect in School 1“ „Effect in School 2“ n4 n5 Analyst: Pam Figure 12 Use of concepts (INT, GRA, SEL) and higher-order concepts (HOC1, HOC2) that are applied to different sets of objects (τ vs. σ) their grades to some extent. Note that we use p in- stead of c relationships in this display, because the two types of data are empirically related (probably because they both reflect the students’ actual cognitive abilities), but the students’ test results are not the cause of their grades. The figure also tells us that this predictive relation- ship is assumed to be different for students in School 1 as compared to students in School 2. We learn this from comparing coefficients p1 and p2, and from how the two higher-order concepts (HOC1 and HOC2) are named. At this point, a new set of objects (σ) has to be introduced to distinguish the two schools from one another. We also learn that something else (SEL) named “selectivity” is assumed to vary across the same objects (σ), and that this variation causally explains the differ- ence between p1 and p2. Discussion In this paper, we presented the first version of VAST, the Visual Argument Structure Tool. It provides a set of clear rules by which the ways in which people think and speak about things may be visually organized, in order to help understand those ways better. VAST may be used for constructing new arguments, as well as for analysing, completing, revising and/or (partly) refuting existing ones. The system captures some of the key types of elements that many everyday arguments and scientific theories share by means of a relatively small set of graphical symbols. In our view, its appeal lies in its intuitiveness and relative economy, in its capacity to account for any degree of specification or relative fuzzi- ness, and in its applicability to basically any content do- main. In the next section, we give a brief overview of the system’s possible uses. Depending on the material at hand, one or several of the following goals may play a more prominent role in a VAST analysis: (a) afford comprehensiveness, (b) explicate premises in terms of what IS the case and what OUGHT to be case, (c) clar- ify the views that different (groups of) people have, in- cluding areas of (dis-)agreement, (d) identify areas of under-specification, inconsistency or outright contradic- tion, (e) deduce defensible conclusions. Potential Uses of the System Theory Specification. There is no shortage of complaints that the mainly narrative theories which are so common in the human- ities and in psychology are of limited value because of their relative fuzziness and under-specification. At the same time, proposals as to how this situation may be improved are largely lacking, and the existing ones are often relatively unspecific themselves. We think that VAST may offer a solution to this problem, for two rea- sons: First, a VAST analysis may help pinpoint those parts of a narrative theory that may and should be better specified, and then aid in the specification process. We consider this the preferable approach compared to the alternative of rejecting narrative theories altogether. By specifying a theory better, its “empirical content” (em- pirischer Gehalt; Glöckner and Betsch, 2011; Popper, 2002, page 96) and thus ultimately its utility will be improved (e.g., it will become easier to refute). Second, VAST allows for accommodating any level of fuzziness that seems acceptable or unavoidable at present. This very much aligns with the idea of scien- tific theory-development as an incremental process of gradually increasing specificity. Furthermore, by being 13 able to incorporate natural-language components of an argument, a VAST display may help bridge the chasm that exists between the humanities and the “harder” sciences, with psychology dangling somewhere in be- tween. In the Appendix we provide a somewhat more complex example, showcasing an attempt to clarify the meaning of a short theory paragraph from a research paper. Nomenclature Issues. Psychology in particular has long suffered — and con- tinues to suffer — from significant jingle- and jangle- problems (Block, 1995): To this day, psychologists of- ten use different words to denote the same thing, or the same words to denote different things. Both practices are at odds with scientific ideals of efficiency and parsi- mony. VAST may be used to help improve on the present situation quite a bit by making visible (at a glance), (a) which terms are used, (b) by whom, (c) to denote what (including the relationships among the concepts that the terms refer to). The according displays will almost certainly involve naming and implication rela- tionships as well as the perspective element. For exam- ple: Is the thing that is called “narcissism” by author A the same as the thing that is called “narcissism” by author B (e.g., in terms of its assumed or shown re- lationships with other concepts)? To what extent are “arrogance”, “dominance”, and “self-enhancement” just different words for the same thing, and is that the same thing that is also called “narcissism” by some scientists? And so on. Facilitating Scientific Discourse. Through its perspective element, VAST analyses may very well be used to account for the inherently social nature of all scientific discourse (Oreskes, 2020), as they enable an explicit and comprehensive showcasing of points of convergence or disagreement between the views of different scholars studying the same subject. We assume that scientific debates may become signifi- cantly more efficient when making sources of disagree- ment visible and then working through them, one after another, possibly in an iterative fashion. This way, a VAST analysis may actually serve as a kind of road-map to help guide the scientific process (e.g., in systematic attempts at forming consensus). Peer Review. VAST may also be used as a tool in peer review. This seems particularly promising when reviewing a paper from a research field that the reviewer is not that fa- miliar with. In such cases, it may be helpful to first organise the available information in the paper as to (a) how many relevant concepts there are (b) how they are assumed to relate to one another, (c) how they are named, (d) how they were measured, (e) how well these measurements reflect the assumed relationships between the concepts of interest, and so on. Use as a Tool for Gathering Research Data. The use of VAST may also be helpful when people’s belief systems (e.g., so-called conspiracy theories) are the research domain of interest. For example, how aware are people of the logical (in-)consistencies within their own worldviews? What happens if you make them aware of the existing inconsistencies (e.g., do they add new components post hoc that mitigate them)? Which components of people’s belief systems are particularly hard to change (e.g., the ones that are of key im- portance to several intertwined belief systems)? And: do people find it easier to map arguments they agree with, as compared to arguments with which they do not agree? Explicating the Paths from Premises to Conclusions (and Back). VAST may be used to derive defensible conclusions from a given set of premises, or to elucidate the ways in which a given perceiver seems to draw conclusions from such premises. Likewise, VAST may be used to in- fer the premises upon which some existing set of conclu- sions was built. Often, perspectives may be of particular importance in such analyses. This is because believing different things to be true (IS) or desirable (OUGHT) goes a long way in explaining wildly different conclu- sions (e.g., in terms of how one should act). Finding Common Ground. The potential use of VAST for working toward con- sensual viewpoints is not limited to scientists (see above). We hope that VAST may just as well be used to enable more traceable and rational conversations among proponents of viewpoints that may seem irrecon- cilable at first (e.g., regarding abortion, second amend- ment rights, vaccination etc.). The extent to which this hope is warranted will have to become the subject of future research, however. Teaching Critical Thinking and the Art of Argumenta- tion. VAST may be used as a teaching tool, helping teach- ers explain to students the various ways in which con- cepts may be related to one another, and the important roles that IS and OUGHT statements as well as differ- ent perspectives play in many arguments. Ideally, these 14 things would be taught by way of analysing existing ar- guments together, or by jointly developing new ones (cf. Cullen et al., 2018). Comparison with Related Tools VAST’s intended domain of use overlaps very signif- icantly with those of many other systems, most promi- nently Directed Acyclic Graphs (DAG; Pearl, 1995; Pearl and Mackenzie, 2018; Rohrer, 2018) and Structural Equation Modelling (SEM). Major points of conver- gence with these systems are obviously the display of concepts or variables (“nodes”), and the display of re- lationships between them (“edges”). In SEM, numerical coefficients are often used to express the strengths of re- lationships between variables. A similar route is taken when creating so-called “research maps” summarizing the theoretical assertions and the evidence speaking for or against them, for a given research field (Matiasz et al., 2018). From SEM, VAST has further borrowed the use of “noise arrows” to symbolise additional, unspeci- fied influences. Diverging from SEM conventions, how- ever, VAST uses a different default meaning for the ab- sence of arrows between concepts: Whereas in SEM this usually signals an unrelatedness of variables, in VAST it means “unrelated or not related in ways relevant to the current argument”. This is a somewhat more liberal ap- proach, and more in line with how everyday arguments are structured, according to our experience: When peo- ple do not talk about relationships between concepts, this usually means that they see no reason for doing so, but not necessarily that they assume the respective coefficient to be zero. The coverage of VAST exceeds those of DAGs and SEMs by a wide margin: Like DAGs and SEMs, VAST does cover the (measured or unmeasured) features of objects and the relationships (causation and associa- tion) among those features. Unlike those other two frameworks, however, VAST also covers some more “psychological” relationship types such as naming, con- ceptual implication and reasoning. For this reason, the default coefficient of relationship strength in VAST ex- presses the covariation of two concepts in terms of per- centages of ranges (e.g., how much more will I consider an object to “be a car” if it “has tires?”). VAST also goes beyond the aforementioned systems in that it enables an explicit accounting for assumptions as to how much something IS and OUGHT to be the case, and for dif- ferences between people in regard to such assumptions. All of this is unquestionably of key importance for many everyday arguments. The coverage and methodology of VAST overlaps considerably with tools developed in philosophy, such as MindMup (https://maps.simoncullen.org/), Rea- son!Able (Van Gelder, 2002) and others (for an overview of argument visualisation approaches, see Okada et al., 2014). These other tools usually enable users to zoom in on any parts of a verbal argument and deduce the logical relationships among them. This con- cerns reasons for drawing certain conclusions as well as objections to doing so. In VAST, these are captured us- ing positive or negative type r relationships. Many tools account for premises that may lead to certain conclu- sions either by themselves, or in combination. In VAST, these would be distinguished from one another in terms of separate vs. combined (AND/OR) arrows pointing toward a concept. Several tools also afford the possi- bility of making whole strains of argument (e.g., “X is a reason to believe Y”) the subject of further reasoning (e.g., “Z is a reason not to believe that X is a reason to believe Y”). In VAST, this is captured using higher-order concepts. Variants of IS-statements and quantifications of reasoning strength are also found in some existing tools (e.g., Reason!Able). However, a major difference between these tools and VAST is that the former deal exclusively with relationships of the reasoning (r) type, whereas the latter also accounts for many other possible types of relationships between concepts, while still cap- turing their strengths with the same (default) metric. Limitations and Outlook At this early stage in the development of VAST, it is difficult to predict how eagerly it will be picked up and eventually be used by others. We have spent signifi- cant amounts of time over the course of approximately three years developing, testing, revising and refining the system through numerous iterations, trying to make it work with analyses of diverse sets of examples both from within science and outside of it. We are convinced that the current version does work reasonably well, but we certainly expect additional improvements in the fu- ture. To facilitate these, we encourage our readers to give it a try, to put the system to use on whatever argu- ments they find interesting, and to let us know about the experiences they make. News media articles, statements by political agents (e.g., parties or office-holders), court sentences, advertisements, and of course science texts are all fair game. Based on our own experiences, we predict that, like us, most readers will find this type of analysis intellectually challenging. We hope that the substantial effort that tends to be associated with speci- fying argument structures this way will not deter people from trying. VAST analyses may be real eye-openers in regard to the, well, vast level of complexity that does permeate many arguments but tends to be overlooked when stick- ing to purely narrative ways of formulating them. In https://maps.simoncullen.org/ 15 this regard, a reviewer (Julia Rohrer) alerted us to a po- tential conflict of interest, especially for scientists who consider using the system: Given the current incentive structure in academia, there may be good (i.e. rational) reasons to avoid greater levels of theory specification. In fact, low levels of specification may be selected for be- cause using them is likely to involve a lower risk of be- ing proven wrong. Ill-specified theories will also be less likely to incur strong negative reactions from reviewers and may be easier to “sell” to the public. All of this is true and may in fact explain part of why weak theoriz- ing is so persistent in psychology to this day. We do con- sider it unlikely that the present paper will incite a mass movement of theory specification enthusiasts. However, based on our own experience with the tool, we do be- lieve that for those psychologists who already are gen- uinely interested in improving on the specificity of their (and others’) theorizing, using VAST is definitely worth a try. Considerably greater clarity is usually achieved. At present, VAST merely consists of a set of concep- tual distinctions and related rules for how they should be visualized. As this is the core of the system, complete and satisfactory VAST analyses are already possible us- ing any standard graphics tool, or even just paper and pencil. However, our ultimate goal is to implement the system as a free web resource that will be capable of (a) developing VAST displays by asking users the right questions and (b) checking any given display for consis- tency and completeness. Appendix: A More Complex Example The task at hand is to clarify the meaning of the fol- lowing theory paragraph from the paper by Theves et al. (2020), using VAST. Such a clarification attempt may — and usually does — lead to an identification of ar- eas of underspecification, ambiguity or even contradic- tion. Note, however, that this particular paragraph was picked for no other reason than being a relatively typical example of narrative theorizing in psychology (and be- cause it deals with the subject of “concepts”, which play a key role in VAST). We do not consider this a particu- larly problematic case, but rather just use it to showcase a typical application of VAST. Thousands of other para- graphs may have been used just as well. The paragraph from the (Theves et al., 2020) paper (page 7318) goes like this: "Concepts are organizing structures that define how contents are related to each other and can be used to transfer meaning to novel input (Smith & Medin, 1981; Kemp, 2010). Their formation thus inherently depends on generalization over, and integration of experiences. Thus, a role of the hippocampus in generalization seemed con- siderable due to its roles in binding elements into spatial and episodic context (Davachi et al., 2003; Komorowski et al., 2013; Davachi, 2006; Ranganath, 2010) as well as integration of information over episodes (Schlichting et al., 2015; Davis et al., 2012; Collin et al., 2015; Milivo- jevitch et al., 2015 [...]" Figure 13 shows a VAST display created by Daniel (analyst) to reflect the paragraph above. More specifi- cally, it shows what this analyst thinks the authors (per- spective) of the paragraph are saying (IS), which is ev- erything inside the largest frame. Note that, for sim- plicity, a minor portion at the end of the narrative para- graph was excluded, as indicated by the use of brackets at the end (see above). Daniel used the default mode of VAST in which all concepts and their names are separated from one an- other. He identified 8 relevant concepts to be accounted for. Five of these (C1 to C5) reflect key sentences from the paragraph, as shown in the lower part of the figure where the respective naming relationships are listed. Note that some minor inferences and modifica- tions were necessary to enable each concept name to speak for itself. The other three concepts (L1 to L3) represent the references to the literature that also fea- ture in the paragraph. These are empirical in nature, as indicated by the use of thick black edges on the re- spective frames. They are also assumed to be given, as indicated by the use of IS pentagons. In this first VAST display, Daniel focuses on what he thinks is the reasoning structure in the paragraph. His use of VAST’s default mode along with his decision to set the naming relationships aside allows us to fully con- centrate on this reasoning structure, without being dis- tracted by the actual content of specific concepts. There is no need for indexing the naming relationships in this context, because they will not be individually discussed. Daniel identified seven relevant relationships of the reasoning type (r1 to r7): Because L1 is given (IS), we may believe C1 to be the case, via r1. This is how Daniel interprets the first citation. If C1 is the case, we may also believe C2 to be the case, via r2. This is how Daniel in- terprets the first “thus”. If all of this is the case, we may also believe C3 to be the case, via r3. This is Daniel’s interpretation of the second “thus”. However, this (r3) is only true if C4 and C5 are also the case, via r4 and r5. This latter conditioning is how Daniel interprets the “due to” in the paragraph. Furthermore, L2 is given, which is a reason (via r6; reflecting the second citation) to believe C4 to be the case. Finally, L3 is also given, which is a reason (via r7; reflecting the third citation) to believe C5 to be the case. Note that the analyst had to make numerous auxiliary assumptions about things that were not entirely clear in the narrative paragraph itself: For example, the analyst 16 “[…] a role of the hippocampus in generalization seemed considerable” “[hippocampus’s role] in binding elements into spatial and episodic context” “[hippocampus’s role] in integration of information over episodes” C1 r3r2 Analyst: Daniel IS Authors C2 C3 C1 “Concepts are organizing structures that define how contents are related to each other and can be used to transfer meaning to novel input.” C2 C3 “[Concepts’] formation […] inherently depends on generalization over, and integration of experiences.” C4 C5 “Davachi et al., 2003; Komorowski et al., 2013; Davachi, 2006; Ranganath, 2010“ “Smith and Medin, 1981; Kemp, 2010”L1 L2 C4 C5 r4 r5 r1 n n n n n n n “Schlichting et al., 2015; Davis et al., 2012; Collin et al., 2015; Milivojevitch et al., 2015“ L3 n L1 L3 L2 r6 r7IS IS 1 1 1 Figure 13 Reasoning structure in a theory paragraph by Theves et al. (2020), according to Daniel (analyst) treated all of the papers supposedly supporting a given proposition as a unified whole (i.e., a single concept). Alternatively, he could have used a separate concept for each paper, which would have meant that each paper by itself supports the respective proposition. Also, Daniel treated r4 and r5 as two entirely separate paths by which r3 is supported. Alternatively, he could have interpreted the “as well as” in the respective sentence to mean that only the combination of C4 and C5 is a reason to believe C3, which he would then have had to express using an AND diamond. Needless to say, each of these additional assumptions, and any other element in the display, may be challenged any time — but only if they are made explicit, which is the whole point of VAST analyses. By making specifi- cation gaps and ambiguities explicit this way, a VAST analysis may eventually help drive theory development. As a next step, Daniel decides to zoom-in a bit more on two of the key concepts in the paragraph. The first of these is C1, which he interprets to be mostly about conceptual implications. This is how he interprets the word “are” in this concept’s name. Figure 14 shows the outcome of this zooming-in: Daniel believes the content of C1 and the entire con- tent of the largest frame underneath it to be inter- changeable. To him, these two concepts have the ex- act same meaning, as indicated by his using a bidirec- tional implication arrow with coefficient 1 (signalling identity). What Daniel does here is specify the meaning of a higher-order concept (C1) by breaking it down into four components (C6 to C9) and some relevant relation- ships (i2 to i4) between them: Objects that are consid- ered exemplars of C6 (named “concepts”) are also con- sidered exemplars of C7, C8 and C9. Note that these latter implications use unidirectional arrows only. Note further that all of this is expressly Daniel’s opinion, and not necessarily shared by the authors of the paragraph. This is because only Daniel is named as the analyst in the byline, but the original authors of the paragraph do not appear as perspective-holders anywhere in the fig- ure. The second concept that the analyst chooses to “zoom-in” on is C2. Figure 15 showcases the result. As in the previous analysis, Daniel assumes that the con- tent of C2 and the content of the largest frame under- neath it are mutually interchangeable (i5). Again, he explicates the meaning of a concept (C2) by breaking it down into a few subordinate concepts (C10, C11, and 17 “Organizing structures” “Define how contents are related to each other” “Can be used to transfer meaning to novel input” “Concepts” i2 i3 i4 Analyst: Daniel C6 C7 C9 C8 n n n n C1 “Concepts are organizing structures that define how contents are related to each other and can be used to transfer meaning to novel input.” n i1 1 Figure 14 “Zooming-in” in on concept C1 (ignoring citations) C12) and some relationships among them (this time of the c type, which is how Daniel interprets the word "de- pends" in the text). Daniel’s use of an AND diamond signals that C12 may only be the case if C10 and C11 are both given, but not if only C10 or C11 are given. This is how he interprets the word "and" in the text. Furthermore, the c 0 in- fluence on C12 makes it clear that there are no other causal pathways by which C12 may come about. This is how Daniel interprets the word “inherently" in the text. However, the question mark above the third arrow pointing toward the AND diamond signals that Daniel is not sure whether another influence is needed to bring about C12. This is because the name of concept C2 only seems to say that C10 and C11 are necessary for C12 to happen, but not that they are sufficient. Here, the analysis points to a need for greater specification. Finally, a brief word on the sets of objects that the concepts in the three figures pertain to. These are ob- viously not the same. In Figure 13, the objects are con- ceivable realities: the one that is described as given (in which certain rules apply and certain research papers exist), and possible alternative ones. In Figure 14, the objects are rather ill-defined. It may be useful to think of them as being various kinds of mental phenomena here (e.g., perceptions, memories etc.). In Figure 15, the ob- jects may be thought of as variants of people’s develop- mental trajectories: Only in those in which C10 and C11 (and maybe something else, see question mark) take place, will we also see C12 happening. For simplicity, we did not explicitly account for the different sets of objects in Figures 13, 14, and 15. “[Concepts’] formation […] inherently depends on generalization over, and integration of experiences.” Analyst: Daniel i5 1 “[Concepts’] formation” “Generalization over [experiences]” “Integration of experiences” C2 n AND C11 C10 C12 n n n c 1 c 0 ? Figure 15 “Zooming-in” in on concept C2 (ignoring citations) Author Contact Corresponding Author: Daniel Leising (ORCID: 0000-00001-8503-5840), Technische Universität Dres- den, daniel.leising@tu-dresden.de Conflict of Interest and Funding The authors state that they have no conflict of inter- est to declare. This project was not connected to any particular source of third-party funding. Author Contributions Daniel Leising: Conceptualization, Methodology, Vi- sualization, Writing — Original Draft, Writing — Re- view & Editing; Oliver Grenke: Conceptualization, Methodology, Visualization, Writing — Review & Edit- ing; Marcos Cramer: Conceptualization, Methodology, Visualization, Writing — Review & Editing. Acknowledgments The authors would like to thank Moritz Leistner for help with layouting this paper. Open Science Practices This article is theoretical and as such is not eligible for Open Science badges. The entire editorial process, including the open reviews, is published in the online supplement. 18 References Baroni, P., Gabbay, D., Giacomin, M., & van der Torre, L. (2018). Handbook of formal argumentation. College Publications. Block, J. (1995). A contrarian view of the five-factor ap- proach to personality description. Psychological Bulletin, 117, 187–215. https : / / doi . org / 10 . 1037/0033-2909.117.2.187 Boole, G. (1854). An investigation of the laws of thought: On which are founded the mathematical theo- ries of logic and probabilities (Vol. 2). Walton; Maberly. Borsboom, D., van der Maas, H. L. J., Dalege, J., Kievit, R. A., & Haig, B. D. (2021). Theory construction methodology: A practical framework for build- ing theories in psychology. Perspectives on Psy- chological Science, 16(4), 756–766. https://doi. org/10.1177/1745691620969647 Büning, H. K., & Lettmann, T. (1999). Propositional logic: Deduction and algorithms (Vol. 48). Cam- bridge University Press. Cullen, S., Fan, J., van der Brugge, E., & Elga, A. (2018). Improving analytical reasoning and ar- gument understanding: A quasi-experimental field study of argument visualization. npj Sci- ence of Learning, 3, 21. https : / / doi . org / 10 . 1038/s41539-018-0038-5 Dablander, F. (2020). An introduction to causal infer- ence. https://doi.org/10.31234/osf.io/b3fkw Devezer, B., Navarro, D. J., Vandekerckhove, J., & Buzbas, E. O. (2021). The case for formal methodology in scientific reform. Royal Society Open Science, 8(3), 200805. https : / / doi . org / 10.1098/rsos.200805 Eronen, M. I., & Bringmann, L. F. (2021). The the- ory crisis in psychology: How to move forward. Perspectives on Psychological Science, 16(4), 779–788. https : / / doi . org / 10 . 1177 / 1745691620970586 Frege, G. (1879). Begriffsschrift: Eine der arithmetischen nachgebildete formelsprache des reinen denkens. Verlag von Louis Nebert. Fried, E. I. (2020). Lack of theory building and testing impedes progress in the factor and network lit- erature. Psychological Inquiry, 31(4), 271–288. https : / / doi . org / 10 . 1080 / 1047840X . 2020 . 1853461 Glöckner, A., & Betsch, T. (2011). The empirical content of theories in judgement and decision making: Shortcomings and remedies. Judgement and De- cision Making, 6(8), 771–721. Matiasz, N. J., Wood, J., Doshi, P., Speier, W., Beck- emeyer, B., Wang, W., Hsu, W., & Silva, A. J. (2018). Researchmaps.org for integrat- ing and planning research. PLoS ONE, 13(5), e0195271. https://doi.org/10.1371/journal. pone.0195271 Muthukrishna, M., & Henrich, J. (2019). A problem in theory. Nature Human Behaviour, 3(3), 221– 229. https : / / doi . org / 10 . 1038 / s41562 - 018 - 0522-1 Okada, A., Buckingham Shum, S. J., & Sherborne, T. (2014). Knowledge cartography: Software tools and mapping techniques. Springer. Oreskes, N. (2020). Why trust science? Princeton Uni- versity Press. Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688. https://doi.org/ 10.1093/biomet/82.4.669 Pearl, J., & Mackenzie, D. (2018). The book of why. Basic Books. Popper, K. R. (2002). The logic of scientific discovery. Routledge Classics. Preparata, F. P., & Yeh, R. T. (1972). Continuously val- ued logic. Journal of Computer and System Sci- ences, 6(5), 397–418. https : / / doi . org / 10 . 1016/S0022-0000(72)80011-4 Robinaugh, D. J., Haslbeck, J. M. B., Ryan, O., Fried, E. I., & Waldorp, L. J. (2021). Invisible hands and fine calipers: A call to use formal theory as a toolkit for theory construction. Perspectives on Psychological Science, 16(4), 725–742. https: //doi.org/10.1177/1745691620974697 Rohrer, J. M. (2018). Thinking clearly about correla- tions and causation: Graphical causal models for observational data. Advances in Methods and Practices in Psychological Science, 1(1), 27–42. https://doi.org/10.1177/2515245917745629 Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd (Eds.), Cognition and cat- egorization (pp. 27–48). Lawrence Erlbaum As- sociates. Smaldino, P. (2017). Models are stupid, and we need more of them. In R. R. Vallacher, S. J. Read, & A. Nowak (Eds.), Computational social psy- chology. Routledge. https://doi.org/10.4324/ 9781315173726 Smaldino, P. (2019). Better methods can’t make up for mediocre theory. Nature, 575, 9. https : / / doi . org/10.1038/d41586-019-03350-5 Theves, S., Fernández, G., & Doeller, C. (2020). The hippocampus maps concept space, not feature space. The Journal of Neuroscience, 40(38), 7318–7325. https : / / doi . org / 10 . 1523 / JNEUROSCI.0494-20.2020 https://doi.org/10.1037/0033-2909.117.2.187 https://doi.org/10.1037/0033-2909.117.2.187 https://doi.org/10.1177/1745691620969647 https://doi.org/10.1177/1745691620969647 https://doi.org/10.1038/s41539-018-0038-5 https://doi.org/10.1038/s41539-018-0038-5 https://doi.org/10.31234/osf.io/b3fkw https://doi.org/10.1098/rsos.200805 https://doi.org/10.1098/rsos.200805 https://doi.org/10.1177/1745691620970586 https://doi.org/10.1177/1745691620970586 https://doi.org/10.1080/1047840X.2020.1853461 https://doi.org/10.1080/1047840X.2020.1853461 https://doi.org/10.1371/journal.pone.0195271 https://doi.org/10.1371/journal.pone.0195271 https://doi.org/10.1038/s41562-018-0522-1 https://doi.org/10.1038/s41562-018-0522-1 https://doi.org/10.1093/biomet/82.4.669 https://doi.org/10.1093/biomet/82.4.669 https://doi.org/10.1016/S0022-0000(72)80011-4 https://doi.org/10.1016/S0022-0000(72)80011-4 https://doi.org/10.1177/1745691620974697 https://doi.org/10.1177/1745691620974697 https://doi.org/10.1177/2515245917745629 https://doi.org/10.4324/9781315173726 https://doi.org/10.4324/9781315173726 https://doi.org/10.1038/d41586-019-03350-5 https://doi.org/10.1038/d41586-019-03350-5 https://doi.org/10.1523/JNEUROSCI.0494-20.2020 https://doi.org/10.1523/JNEUROSCI.0494-20.2020 19 Van Gelder, T. (2002). Argument mapping with rea- son! able. The American Philosophical Associa- tion Newsletter on Philosophy and Computers, 2(1), 85–90. Introduction The System Concepts Types of Relationships Between Concepts Relationship Type 1: Naming (n). Relationship Type 2: Conceptual Implication (i). Relationship Type 3: Causation (c). Relationship Type 4: Transformation (t). Relationship Type 5: Prediction (p). Relationship Type 6: Reasoning (r). Additional Relationship Types. Relationship Strength Noise Relationship Direction The IS and OUGHT elements The Perspective Element The Analyst Element Higher Order Concepts Discussion Potential Uses of the System Theory Specification. Nomenclature Issues. Facilitating Scientific Discourse. Peer Review. Use as a Tool for Gathering Research Data. Explicating the Paths from Premises to Conclusions (and Back). Finding Common Ground. Teaching Critical Thinking and the Art of Argumentation. Comparison with Related Tools Limitations and Outlook Appendix: A More Complex Example Author Contact Conflict of Interest and Funding Author Contributions Acknowledgments Open Science Practices