Associating With Art: A Network Model of Aesthetic Effects

Aesthetic experiences are complex experiences that unfold over time and assume a strong relation between features in artworks and their effects on the beholder. But how does this actually happen? A key concept is aesthetic effect which, since the 18th century, has been used in art literature to explain responses to artworks (Brinkmann et al., 2018; Schnackertz, 2010). The assumption is that works of art produce specific aesthetic effects in viewers that are caused by single elements, such as a red causing a sensation of warmth, or a horizontal line causing a quiet sensation. For example, Franz Marc described colors in a letter to August Macke as follows: “Blue is the male principle, astringent and spiritual. Yellow is the female principle, gentle, gay and sensual. Red is matter, brutal and heavy and always the color to be opposed and overcome by the other two” (1964, p. 28). Similar descriptions can be found for lines. Kandinsky, for example, postulated: “The same thing can with complete justification be said about the horizontal and the vertical - low and high. The former is lying, the latter is standing, walking, moving about, finally climbing upward. Supporting - growing. Passive - active. Relatively: feminine - masculine” (Kandinsky, 1947, p. 63).

Note that aesthetic effects are used to describe how the artwork appears to the viewer rather than the viewers direct experience. The viewer does not feel “gay” when looking at yellow, nor are they “lying” when looking at horizontal lines. Rather, the yellow appears “gay” to the viewer, and, similarly, horizontal lines are seen by the viewer as “lying”. Therefore, these aesthetic effects can best be translated as associations in psychological nomenclature. Aesthetic effects are often described as co-occurring—multiple effects such as warmth and quiet are simultaneously caused by the same artwork—but their connection to each other is rarely made explicit.

In this paper, we propose a novel way to conceptualize these effects by explicitly connecting them to each other. In stark contrast to prior work, we conceive these effects as an associative process of direct relationships between aesthetic effects by utilizing a network approach.

We first describe the theoretical model employed by previous research to highlight the contrast with our proposition of an associative process. Second, we introduce our theoretical model, the aesthetic effects network (AEN), and discuss the results of an empirical study (Study 1) that was conducted to explore what information can be gained from taking this approach. Third, we discuss the results of a pre-registered follow-up study (Study 2) that provides evidence for the AEN over other explanations.

A major body of previous empirical work on aesthetic effects consists of artistic research (Borgdorff, 2009, 2012)—in which artists experiment with their art in an attempt to create a specific effect—rather than scientific research (Brinkmann et al., 2018). The few scientific studies examining associations with art, such as Takahashi (1995), relied on Osgood et al.’s EPA model (Osgood et al., 1957), which was developed to capture semantic space or meaning in three latent factors: Evaluation, Potency, and Activity.¹ To measure meaning, Osgood et al. proposed a semantic differential (or bipolar scale), which uses two opposing adjectives such as happy–sad, active–passive, or strong–weak.

For aesthetics that means that, when people look at an artwork, they will perceive a certain level of Activity in the artwork, which will consequently lead to a related score in the corresponding indicator. For example, if the Activity is high, a person will rate the painting as lively rather than still. In this model, relationships between aesthetic effects are due to a common cause: if active and lively are correlated this is because they are both caused by the latent factor of Activity.

In our view, relationships between aesthetic effects should be seen as direct associations between one aesthetic effect and another; this presupposes a level of causality where the aesthetic effects can cause each other. Thus, the main difference between the common cause perspective (EPA model) and our proposed AEN model (see 1.2 for details) is the conceptualization and interpretation of relationships between aesthetic effects. In the EPA model, this relation exists because they belong to the same factor. In the AEN, these are interpreted as direct relationships. The conceptual notions behind the EPA and AEN model are illustrated in Figure 1.

Approaches such as the EPA model represent a relationship between artwork and viewer, where the artwork can inspire aesthetic effects in the viewer who is a passive receiver. The AEN model claims a more active role of the viewer. Though the initial stimulation comes from the artwork, the viewer can freely associate throughout the AEN: this process constitutes an aesthetic experience.² This is analogous to semantic networks such as word associations that have been widely described as network structures (De Deyne & Storms, 2008; Steyvers & Tenenbaum, 2005), where the activation of one word (e.g., “love”) leads to the activation of another word (e.g., “heart”). These networks have been used to predict performance in memory retrieval tasks and explain various priming and interference phenomena (Collins & Loftus, 1975; Nelson et al., 1998). One especially relevant concept in this context is that of “spreading activation” (Anderson, 1983) which holds that, as one node in the network is “activated”, this activation can then spread throughout the network. We propose that a similar process would underly the AEN. Note that a full investigation on models from cognitive psychology and how they relate to the topic at hand is outside the scope of the current paper. As a first step, we decided to exclude this since we are already aiming to bridge art history, empirical aesthetics, and network modeling. Nonetheless, as this approach develops, it should consider the work done in cognitive psychology in more detail.

Finally, note that the main difference between the EPA and the AEN model is theoretical (see also our rationale for Study 2 below) rather than a distinction between factor analytic and network analysis. Our aim with the current paper is to introduce this new theoretical point of view, rather than to introduce a new methodology. As noted above and illustrated in Figure 1, the theoretical difference focuses around the interpretation of the interrelations between aesthetic effects. Finally, we view the current paper as a first step into this new theoretical direction, while at the same time realizing that much additional work is needed to substantiate this theory. Since this perspective on the future is thus also a significant contribution of the current paper, we address this at length in the Limitations and Future Directions section at the end of the paper.

We propose to model the associative process between aesthetic effects as a statistical network. Networks consist of nodes and edges. Nodes represent entities and edges represent connections between these entities. This general concept is illustrated by the notion of “six degrees of separation”, the idea that any two random people on earth are six or fewer acquaintances apart, originally tested empirically by Travers & Milgram (1969). This is an example of a social network where nodes represent people, for example Harry and Sally, and edges represent relationships between them. Thus, if Harry and Sally know each other, there will be an edge connecting the node of Harry to the one of Sally.

In the AEN, aesthetic effects (active, lively, etc.) are represented by nodes. Associations between the aesthetic effects are represented as edges. Thus, if active is associated with lively, there will be an edge connecting the nodes. What “associated” in the previous sentence means is not trivial: in a social network, we can simply observe if people are friends (e.g., on Facebook); in the social sciences, we require psychometric models to estimate such relations.

We propose that in the AEN, the edges will not be binary (i.e., either present or absent) but are more appropriately modelled as continuous, representing the strength of associations. In addition, active and lively can be positively associated (i.e., the more active, the more lively) or negatively associated (i.e., the more active, the less lively). In our case, we conceptualize these edges as partial correlations because we are interested in the shared unique relations among aesthetic effects partialling out all other aesthetic effects. This means that the variables are dependent conditional on all other items in the network (for an overview of these models, see Epskamp & Fried, 2018).

Aesthetic effects occur as a reaction to a stimulus much like attitude networks (Dalege et al., 2016) that represent attitudes towards a specific object or person—without president Trump, there is no attitude towards president Trump. We term such networks stimulus-bound since they only exist as a reaction towards a stimulus, in contrast to e.g., social networks which are not stimulus-bound. Although separating these out may in some cases be difficult, the main difference is that stimulus-bound networks require elicitation by the stimulus to be observable while non stimulus-bound networks do not.

The stimulus in attitude or aesthetic networks exists outside of the network itself in the external field (Borsboom, 2017), denoting factors that can influence nodes inside the network from the outside (cf. Figure 1). Thus, the stimulus may influence one (or multiple) nodes in the network but is not itself a part of the network and, therefore, not represented in the network by a node. In the case of stimulus-bound networks, we propose that the stimulus would be seen as the common cause—as such, the full model can be conceptualized as a hybrid model (Fried & Cramer, 2017) which is a mixture of a network model and a common cause model—meaning that the stimulus activates certain (not necessarily all) nodes in the network. For example, if Sally perceives a painting, this painting can cause the association of active in her. Since the association active is a node in the network, the activation can then spread from there throughout the network (see Figure 1). This also opens up the possibility for a network approach to individual differences in aesthetic experience. We will return to this issue in the Limitations and Future Directions section.

Here we report the results of an empirical study to put forth a preliminary account of the structure of the relationships between different aesthetic effects.

We surveyed 255 voluntary participants (92 male, 160 female, 2 identifying as other, 1 unreported, mean age = 20.73, SD = 3.73) in a first-year psychology lecture. We collected a convenience sample of participants with the aim of recruiting as many participants as possible in a single session. All experiments of the present paper were carried out in German, in accordance with the Declaration of Helsinki, and after approval by the local ethical committee of the University of Vienna.

To measure aesthetic effects, we included 14 items. Each item was a semantic differential, where each aesthetic effect was represented by opposite pairs as poles of the scale, for example: warm–cold. The choice of the semantic differential format and of our specific word pairs was based on a recent historical analysis of the most frequently used terms for aesthetic effects in art literature (Brinkmann et al., 2018). We used these pairs of terms: negative–positive, passive–active, lively–still, happy–sad, aggressive–peaceful, soft–hard, warm–cold, heavy–light, smooth–rough, bodily–spiritual, masculine–feminine, intrusive–cautious.³ In addition, we included dislike–like and uninteresting–interesting because they are outcome variables of aesthetic experience that are widely considered as relevant and important by psychologists as shown by, for example, the influential Pleasure-Interest Model of Aesthetic Liking by Graf & Landwehr (2015) and the seminal work by Daniel Berlyne (1974).

In each case, the left term was represented by 1 and the right term by 7. Left term and right term are here used to refer to the opposite sides of the dimension such as “warm-cold”, in this case warm would be referred to by 1 and cold by 7. The specific wording of questions for each scale was as follows, “This image appears … warm/cold” (In German: “Dieses Bild wirkt … warm/kalt”). For liking the question was rephrased to “I like/I do not like this image” (in German: “Dieses Bild … Gefällt mir/Gefällt mir nicht”) in order to stay grammatically correct (please note that in German the grammar between liking and the other scales is much more similar than in English).

As stimuli we used 8 abstract artworks by 5 different artists. Image choice was based on Brinkmann et al. (2018) and Specker et al. (2020); all images represented work by artists that can be directly related to art historical theories on aesthetic effects. All image information and their shorthand titles can be found in Table 1. For ease of reading we chose shorthand titles for each of them (see third column) to be used throughout the text and in all figures and tables. We chose a relatively homogenous set of artworks, because we aimed to average across artworks to derive a general network structure. Data on no other artworks was collected, i.e., we did not perform a sub-selection of specific artworks for the purpose of this paper.

The study took place during a first-year psychology lecture. All images were projected against a black background by use of a projector. Thus, image presentation was consistent across participants. Participants were instructed to first look at each image, let it make an impression on them, and then to rate the image on the 14 aesthetic effects within a total viewing time of 75 seconds for each image. After each image, a blank black screen was shown for 5 seconds before moving on to the next image. Responses could be given during the entire viewing time (75 + 5 sec). Ratings could be made either with pen and paper or a digital answering option (tablet, PC, and phone). Pen and paper responding (N = 127) was partly randomized: each participant had the same question order for each image, but the question order was different for each image. For digital responding (N = 128), the question order was random for both images and participants. All scales were presented on one page.

We employed a 3-step analytic strategy similar to Rhemtulla et al. (2016; for a similar, more recent example see Fried et al., 2018). In their study, Rhemtulla et al. (2016) used network analysis to compare the interrelation of symptoms in six different datasets of patients with different substance abuse disorders. To do this, they first estimated a symptom network for each substance abuse disorder dataset. Then they estimated a global network across all datasets to understand what a general network of substance abuse symptoms would look like. Finally, they estimated a variability network to assess the variability of the edges across the six individual substance abuse disorder networks. Though we are not at all examining mental disorders, our aims in the current study are comparable: we want to compare the interrelation of aesthetic effects for multiple artworks and estimate a global network of aesthetic effects. A variability analysis then adds to our understanding of how associations between the aesthetic effects differ depending on which artwork people are viewing. Therefore, we followed the approach of Rhemtulla et al. (2016) and performed three analyses. In contrast to Rhemtulla et al. (2016), and based on Fried et al. (2018), we only interpret the centrality in terms of expected influence, because strength takes absolute values and thus does not take negative edges into account, and betweenness and closeness centrality are often unreliable (e.g., Epskamp et al., 2018). How to interpret centrality statistics (i.e. inference) is being actively debated in the literature (Bringmann et al., 2019; Christensen & Golino, 2020; Hallquist et al., 2019). Here, we use the expected influence statistic, since it provides a measure of how connected one node is to the network. Our interpretation is that more connected nodes are more likely to influence other nodes in the network (as is assumed by accounts of spreading activation, discussed above). This inference is based on our theoretical views, and requires further corroboration before any strong conclusions can be drawn. In addition, we estimated the accuracy and stability for the individual networks (following Epskamp et al., 2018)—corresponding plots can be found in the Supplementary Material. Note that these procedures cannot be applied to the global network as the global network is not estimated in and of itself but is an average over the estimates of the individual networks. All centrality measures were standardized. All analyses were performed in R using the “qgraph” package (Epskamp et al., 2012) implementing a Gaussian graphical model (GGM), a network in which edges connecting aesthetic effects represent estimates of partial correlations, including a graphical lasso regularization. In the GGM, edges can be understood as conditional dependence relations among aesthetic effects: If two aesthetic effects are connected in the resulting graph, they are dependent after controlling for all other symptoms. If no edge emerges, aesthetic effects are conditionally independent. Regularization involves estimating a statistical model with an extra penalty for model complexity (Epskamp & Fried, 2018). This procedure leads to networks that are sparse and constrain many of the small coefficients to zero. This means that edges that are likely to be spurious are removed from the model, leading to networks that are simpler to interpret. Using the graphical lasso to estimate a GGM improves network estimates and leads to a sparse network that describes the data parsimoniously. We used the method of regularization outlined in Epskamp & Fried (2018), using the default setting of the gamma parameter of 0.5. For the global network, we then averaged each edge across the eight separate artwork networks. The GGM with regularization represents the current state-of-the-art for ordinal or continuous data. Note that, while the model aims to map conditional dependence relations, causal claims do not follow from the statistical model, which is why Figure 2 (see below) does not include arrows. The final dataset used for the network analyses and the corresponding code can be found on the Open Science Framework (OSF): https://osf.io/zqxbm/.

We start with the interpretation of the global network of aesthetic effects (averaged across all 8 individual networks), shown in Figure 2. The corresponding centrality plot is shown in Figure 3. The individual networks and their comparison follow afterwards.

The general network of aesthetic effects represents a preliminary account of the general structure of relationships between aesthetic effects. As Figure 2 shows, the estimated network was interconnected with strong positive edges between the nodes like and interesting (.27), light and cautious (.24), positive and interesting (.19), sad and rough (.19), hard and cold (.18), and positive and peaceful (.17). The network has strong negative edges between the nodes peaceful and rough (-.24), cold and light (-.17), hard and light (-.15), and positive and sad (-.15).

With regard to centrality as measured by expected influence, interesting has the highest expected influence in the network followed by positive, cautious, and still. This is plotted in Figure 3. Notably, rough has a very low level of expected influence and spiritual falls somewhere in between.

An alternative approach to estimating a global network would be to average the raw data across artworks (i.e., creating a mean score for each aesthetic effect) and estimate a network from these means. To estimate the robustness of our results, we also estimated the global network this way, the corresponding figures can be found on the OSF: https://osf.io/zqxbm/. We measured the similarity of the two global networks by correlating the adjacency matrices r = .83; differences were mainly due to the second method leading to a sparser network due to regularization.⁴ In addition, the correlation between the centrality of the two networks showed high similarity (r = .94). We conclude that results are reasonably robust to the different estimation methods.

Since the network structure may vary depending on the artwork, we also estimated the aesthetic effects network per artwork, shown in Figure 4. The right panel of Figure 3 represents the corresponding centrality plot. Though we aimed to include homogenous artworks on the assumption that the artworks would be more comparable, we could not a priori know to what extent this assumption was valid. As such, an analysis of the individual artwork networks provides more insight in this direction as well as providing indications on how much (even small) differences between artworks impact the estimated network structure. As noted, we also assessed the accuracy and stability of the resulting networks, following Epskamp et al. (2018), with 2000 samples (corresponding plots can be found in the Supplementary Material). While no clear guidelines exist, based on the prior literature, we consider the networks to be estimated with moderate stability. 95% CIs around edge weights are not small, and point estimates of edge weights hence need to be interpreted with some care. However, the edge weights difference test shows that many edge weights differ significantly from each other for the majority of artworks, indicating that the stronger edges can be meaningfully distinguished from the weaker edges.

Figure 4 shows considerable similarities across the networks. For example, the positive association between interesting and like is present in all networks, and the relation is relatively strong in a consistent way. In addition, only the Winter network (bottom row, 4^th network) includes an association between active and cautious, and the relation is weak. Several differences also emerge. For instance, both artworks by Paul Klee (top row, 3^rd and 4^th network) have a strong negative association between cold and hard, whereas the other networks feature strong positive associations between these aesthetic effects. From the centrality plot (Figure 3, right panel), we can see that there is variability in expected influence depending on artwork. For example, light has a relatively high expected influence for Mortensen Pink and Klee Red compared to the other artworks. There is also some consistency across artworks, for example, interesting and positive have high expected influence for all artworks.

As a next step to investigate cross-network variability in more detail, we derived an index of network similarity by correlating adjacency matrices and centrality across artwork networks (Table 2). The three lowest correlations (between r = .51 and r = .58) of the adjacency matrices correlations are moderate, all other correlations indicate strong similarities. The centrality correlations show a similar pattern, with 5 correlations in the moderate range, and all other correlations in the strong range. To put this into perspective: The highest correlation reported by Rhemtulla et al. (2016), who compared networks of substance abuse symptoms across samples, was .64, with most other correlations in the .20-.50 range.

As a final step, we assessed the variability of edge weights across networks. Figure 5 represents a network in which edges represent the variability in connection strength across the artworks rather than average connection strength. The network is scaled to a standard deviation of .15. In this network, the greatest variability is found between the associations of cold and hard (.14). Another relatively variable association is between cold and light (.12). Looking at Figure 4, we can see that these differences between the association with cold and hard may well be driven by the two artworks by Paul Klee. As noted, they include a negative association between these aesthetic effects while the other artworks include a positive association. Together these results indicate that there is some variation in the associations between the aesthetic effects dependent on artwork—however, aesthetic effects networks seem to be relatively similar in their structure.

These results allowed us to formulate a preliminary account of the general structure of relationships between aesthetic effects. From this general structure, hypotheses can be derived on how aesthetic effects relate to each other. For example, participants in our study who rate an artwork as happy are likely to also rate that artwork as positive, feminine, and smooth since these nodes are strongly connected. Though some of these associations may feel like common sense, recent empirical work has shown that people have a relatively low level of agreement on the aesthetic effects they attribute to artworks (Specker et al., 2020). In addition, we found that (in order) interesting, positive, cautious, and still have the highest expected influence making these nodes more likely to spread activation through the network. The expected influence of interesting and positive was stable across networks, whereas cautious and still varied more for the individual artwork networks. Hence, if the aim is to manipulate the network, one of these four nodes would be a logical candidate. Furthermore, when considering a real-life situation, the fact that interesting had the highest expected influence may suggest that this could be the starting point of an aesthetic experience: We may first need to find an artwork interesting to then start the associative process.

A limitation to Study 1 is that different theoretical models can give rise to the same correlations among items (Fried & Cramer, 2017; Kruis & Maris, 2016; Van Der Maas et al., 2006). It is not possible to figure out which model underlies the data based on estimating models, but it is possible to distinguish between the models by experimental manipulation.

If the underlying model is a network model, changing one node in the network (e.g., making an artwork appear more positive) should lead to a change in other nodes due to their interrelations. In contrast, if the underlying model is a common cause model (like the EPA), this should not happen due to the causal direction the model proposes: Evaluation, in the EPA model, is the shared causal origin for both positive and peaceful.⁵ Thus, making an artwork appear more positive should not change peaceful (or another node) because the underlying cause (Evaluation) has not been changed.

This is perhaps most easily understood when drawing a parallel with illnesses such as measles. If a patient has measles, this will lead to corresponding symptoms such as fever and Koplik’s spots. In this situation, symptoms are passive indicators of the underlying disease, similar to the EPA model. Treating the symptoms will alleviate suffering, but it will not cure the disease: a cure has to target the disease itself, not its indicators. The same applies to our case: Changing the rating of positivity should not lead to a change in Evaluation and thus should not influence how the artwork is rated on the other indicators. Thus, a study that changes one node in the network could differentiate between a common cause and a network model. If other nodes change as a consequence of this manipulation, this would provide evidence for a network model; if they do not, this would provide evidence for a common cause model.

In order to test this, one needs a target node to manipulate. The AEN itself does not give an account of how the stimulus activates the network, it only postulates that the stimulus can activate the network. However, one can draw from the literature to generate hypotheses on how an artwork may activate the AEN. Since research shows that brightness of colors is generally associated with positivity (Lakens et al., 2012, 2013; Meier et al., 2004, 2015; Specker et al., 2018; Specker & Leder, 2018), we decided to manipulate the brightness of artworks as a means to activate the node of positivity. This manipulation is particularly suitable since there are indications that the association between brightness and positivity applies to the general brightness level of complex images (Lakens et al., 2013) as well as that it is universal (Specker et al., 2018), and positivity had a high expected influence in Study 1. This means that the manipulation should be applicable to artworks and be relatively insensitive to individual differences. From our AEN model, this manipulation should lead the node of positivity to be activated, and from this node the activation can spread throughout the network. This is a general hypothesis that assumes that if positivity changes, it can have an effect on the network as a whole. This would be strong evidence for the AEN model, because common cause models (like the EPA) exclude direct influences of positivity on any of the other nodes.

Nonetheless, based on the general network structure identified in Study 1 (Figure 2), we also formulated specific hypotheses as to which nodes were most likely to be affected by our manipulation based on the strength of their connection to positivity. Notably, this manipulation should lead to the artwork being seen as peaceful, smooth, and happy. These are direct effects, but we also assume indirect effects that are caused not by direct connections with positivity but are a result of a longer causal chain. Thus, we predicted that still, peaceful, liking, and interest would go up (lighter images are more still) and that sad, rough, and cautious would go down (lighter images are less sad). We pre-registered our study: https://osf.io/7upyh/register/5771ca429ad5a1020de2872e. To ensure that our manipulation would be effective, we first conducted a pilot-study to test if the brightness–positivity association would translate to artworks.

We adapted the methods of Specker et al. (2018) and Specker & Leder (2018), and used an Implicit Association Test (IAT) as well as an explicit rating task that were counterbalanced. Participants saw dark and bright versions of the artworks used in Study 1. These versions were created by use of GIMP software making them either 30% darker or 30% brighter. We had 27 participants⁶—mean age = 22.93 (SD = 6.94), 18 female, none identifying as other—all students of the University of Vienna who received course credit for their participation. All participants participated voluntarily and completed the experiment in German. Results indicated a strong implicit association between brightness and positivity, t(26) = 13.08, p < .001, d = 2.52, 95% CI:[1.73,3.29], as well as a strong explicit association between brightness and positivity: t(26) = 4.994, p < .001, d = .96, 95% CI:[.49,1.41]. Given the guidelines by Cohen (1988) and also more recent suggestions (e.g., Sawilowsky, 2009), both effects can be considered as large (>.80) and the implicit effect as “huge” (>2.00), indicating that the effects should be generalizable to artworks and that our manipulation should be effective. The dataset and analysis syntax can be found on the OSF: https://osf.io/3n8zy/.⁷

Our method for Study 2 was similar to Study 1. Participants saw eight artworks and rated each artwork on 14 scales. However, in this study, there were three versions (control, dark, and bright) of each artwork. Dark and bright versions were the same as in the pilot-study. Control versions represent “original” versions of the artworks, meaning the same version as in Study 1. Thus “control version” refers to the fact that these images were not manipulated. Since these images are all reproductions of artworks and thus not strictly “originals” we use “control version” throughout the text to refer to these images. From the association between brightness and positivity it should be the case that: bright version > control version > dark version for ratings of positivity which would constitute a manipulation check. If this holds, one can then analyze whether the manipulation had the expected effect.

To make the variation in brightness salient, we wanted to ensure that each participant saw control, dark, and bright versions of artworks. At the same time, to cancel out effects of repeated exposure, we wanted each participant to see each artwork in only one version. We created three groups which differed in which version of the artwork they saw (Table 3). These groups were not used as a basis for analysis. The experimental script, final dataset, and analysis scripts are made available on the OSF: https://osf.io/3n8zy/.

We tested 133 participants (22 male, 111 female, none identifying as other; mean age = 20.41, SD = 2.60). We collected a sample from the same participant pool as Study 1 in order to have comparability across the two studies. All participants participated voluntarily, completed the experiment in German, and were students of the University of Vienna who received course credit for their participation.

As in Study 1, participants saw 8 artworks which they rated on 14 scales. Different from Study 1, participation took place in the lab, the study was self-paced, and presentation order of the scales was randomized. Upon arrival, participants read and signed the informed consent form, and afterwards completed the rating task.

To analyze our results we first did a manipulation check to assess if our manipulation was effective using a one-way ANOVA with version (bright, dark, control) as an within-subjects independent variable and positivity as a dependent variable. If our manipulation was effective it should be that the brighter an image, the more positive it is rated (so: bright version > control version > dark version). This is exactly what we found, F(2,264) = 3.77, p = .024, ηG² = .018, representing a small effect based on Cohen (1988), illustrated in Figure 6. It has to be noted that the effect was smaller than anticipated based on the pilot-study (d _pilot study = .96). This is most likely due to the fact that in the pilot-study participants saw two versions of the same image, thus leading to a more direct comparison based on the brightness feature (making brightness more salient), whereas in this study participants only saw one version of each image.

Second, we conducted a MANOVA with version (bright, dark, control) as a within-subjects independent variable and all aesthetics effects (with the exception of positive) as dependent variables to test whether the manipulation changed how the aesthetic effects were rated. The MANOVA was significant, F(2, 264) = 3.55, p < .001, η² = .154, representing a large effect based on Cohen (1988), which we interpret as evidence for the AEN over the EPA model.

To test our specific hypotheses we ran a one-way ANOVAs with version (bright, dark, control) as an within-subjects independent variable for each dependent variable seperately. All ANOVA results and corresponding means and standard deviations are reported in Table 4. All corresponding figures can be found on the OSF: https://osf.io/3n8zy/.

As predicted, we found that darker images were significantly perceived as sadder, however, we did not find significant effects for liking, interest, still, and peaceful.

One thing is clear from these results: they are not in accordance with a common cause perspective which would be unable to explain the significant MANOVA. Nonetheless, these results do bring up new questions for the AEN.

A limitation to Study 2 is that it operates on simple principles, including selective influence: the idea that one thing would only influence one other thing (rather than multiple other things). In our case, a change in brightness is assumed to directly influence positivity, and only positivity. This assumption can rightfully be questioned, specifically in the case of the found significant effect on heavy–light. It would be possible that this effect was an indirect effect of positivity, however, our manipulation could also have influenced the “weight” node directly, given that the literature on cross-modal correspondences shows an association between brightness and weight (Alexander & Shansky, 1976; De Camp, 1917; Walker et al., 2010; Wright, 1962).

Nonetheless, though a direct influence of brightness on more than just positivity brings up new questions with regard to the AEN, this seems an even more difficult phenomenon for a common cause perspective. Though it would be possible to introduce a new common cause model where brightness of the artwork directly influences all three EPA factors or solely the Evaluation factor, this model would need to be able to explain the results found for specific aesthetic effects. In this model, it seems even more puzzling why brightness would influence happy-sad (and positive-negative) but none of the other terms associated with the Evaluation factor. If there is a direct effect of brightness on Evaluation, then all related specific terms should be influenced. Similarly, the existence of an underlying causal latent factor becomes rather questionable when its indicators can be influenced independently from each other. In sum, it would be possible to create complex factor models that potentially could also explain our results; however, according to Occam’s razor, in competitions of equal probability, the simpler explanation should be given priority. Note that we do not mean to imply here that network models/theory are/is sparse in and of itself, but rather that, given that any kind of network model could explain this phenomenon—whereas only a very specific complex factor model could—it is a sparse(r) explanation for this specific phenomenon.

With regard to our predictions for specific terms, the unexpected results can be explained in several ways; however, the simplest explanation may be that of Simpson’s paradox: a statistical relation exists on one level (e.g., between-persons), but the opposite/no relation exists on another level (e.g., within-person). The AEN (Figure 1) is constructed on the between-person level, contrasting the within-person MANOVA. While this does not influence the main theoretical scope of the AEN—that change in one node should lead to a change in another node—it could influence the predictions based on the specific network structure of the AEN. Though Simpson’s paradox can generally play a role in psychological phenomena, a difference between the between-person and within-person level seems especially plausible for aesthetic phenomena. Common sayings like “beauty is in the eye of the beholder” already hint at between-person differences in aesthetic phenomena. Furthermore, these intuitions have been substantiated empirically (e.g., Leder et al., 2016) and, specifically relevant for our study, recent work (Specker et al., 2020) has shown between-person disagreement with regard to aesthetic effects. At the same time, within-person stability has been observed, even for people with memory impairment (specifically, Alzeihmer’s; Graham et al., 2013). We will discuss how individual differences can be incorporated in a network perspective in more detail in the General Discussion.

We presented a new perspective that changes the way we, at least formally, conceptualize aesthetic effects and, to a larger degree, aesthetic experience. This network perspective introduces a more process-oriented view. While most theoretical models do conceptualize aesthetic experience as a process (Leder et al., 2004; Locher et al., 2010; Pelowski et al., 2017; Tinio, 2013), they generally include a large amount of potentially relevant variables, often with a lack of clear hypotheses regarding the interrelation of the different sub-processes and variables involved. With the AEN, we provide an explicit account of a specific cognitive process involved in aesthetic experience.

Note that, therefore, the AEN supplements previous models rather than replaces them. The aim is to substantiate existing models by providing concrete models for variables, processes, and interrelations of these models. As such, even though the current study focuses on only one cognitive process, this new approach can be applied to aesthetic experience in general. When taking this wider perspective, future work could model other variables of interest, such as relevant individual differences (e.g., art interest) or emotions.

Finally, the AEN offers a new perspective. We believe that this perspective—refocusing on understanding the processes underlying aesthetic experience—can lead to valuable insights into how people experience art. However, the AEN is very much a theory under development. Study 1 aimed to propose a preliminary structure that was consequently tested in Study 2, and results of Study 2 indicate the need to revise and update the theory further; though the general idea of an underlying network was supported, the specific network structure was not. Theory formation is a long process, and we hope this initial report provides impetus for future research. Below, we consider the limitations of the current work and possible future directions in more detail.

As noted in the introduction, the way a stimulus influences the network may vary across people; in addition, the network structure may differ across people (both are illustrated in Figure 7). These differences can potentially explain—or, at least, help us better understand—why people have different experiences when looking at the same artwork, as well as to investigate why people have similar experiences when looking at different artworks. These questions are central if we want to understand how people experience art.

Differences in general structure may occur due to differences in experience (e.g., differences in art interest or knowledge), see below. Similarly, differences in the way a stimulus influences the network may occur due to which parts of the artwork the viewer pays attention to. For example, using eye-tracking, Pihko et al. (2011) showed that art experts had different viewing strategies (as reflected by target, location, and path of fixations) leading them to pay attention to different aspects of the paintings than art novices.

Theoretically, the AEN can be connected to other theoretical models mainly proposed within research on personality—specifically the cognitive-affective personality system (CAPS; Mischel & Shoda, 1995) and other models conceptualizing personality as a dynamic system (Beck & Jackson, 2020; Cramer et al., 2012). What all these approaches share is that they propose to view personality as a dynamic system in order to understand similar questions as to interindividual (in our case, “why do different people have different experiences with the same artwork?”) as well as intra-individual differences (“why do people have similar experiences with different artworks?”, or similarly, “why does the same person have different experiences with different artworks?”). Whereas in personality research there is an interaction between the person and the situation, in art we deal with an interaction between the person and the stimulus. Similar to the approach in personality, what this requires is to look at the person, the stimulus, and their interaction. As Mischel & Shoda (1995) stated: “The theory views the person not as reacting passively to situations nor as generating behavior impervious to their subtle features, but as active and goal-directed, constructing plans and self-generated changes, and in part creating the situations themselves” (p.252). This situation is analogous to what we have in our paper so far called the “active role of the viewer” in the AEN. To give a concrete art specific example, the Pihko et al. (2011) study cited above, indicated that people have different viewing strategies: By paying attention to different aspects of an artwork viewers can, in this way, actively influence their perception and experience of artworks.

Below, we describe some empirical approaches that can be taken in this direction (i.e., group-based comparisons, mixture-network modelling, latent class analysis). Note that for the AEN to progress as theory, it would be necessary to formulate a comprehensive theoretical approach similar to the approaches in personality psychology noted above (Beck & Jackson, 2020; Cramer et al., 2012; Mischel & Shoda, 1995).

Currently, structural differences can be investigated by looking at differences on a group level. To give some examples: art experience differs based on expertise (Chamberlain & Wagemans, 2015; Chirumbolo et al., 2014; Leder et al., 2014; Winston & Cupchick, 1992), thus one could investigate whether art experts (as a group) have a more strongly connected network than lay people (as a group). Based on findings related to art styles (Augustin & Leder, 2006; Leder et al., 2012), one could also compare a network for abstract art with one for figurative art.

Both of these approaches could provide new and interesting information that cannot be gained from methods used so far. However, both require averaging over data in order to generate a network. This is already highly relevant in non-stimulus-bound cross-sectional data where networks are created by averaging over people. In our stimulus-bound case, this becomes even more relevant because we average not only over people, but also over stimuli.

What in the future could be promising is mixture network modeling. Within a mixed model, averaging is not required since both the between and the within-subjects effects of independent variables are modelled at the same time (Baayen et al., 2008). Because these models model random error at all levels of analysis (Nezlek, 2001), they can account for dependencies such as responses to stimuli being dependent on individual participants (i.e., there is a person by stimulus interaction). In fact, recent studies within empirical aesthetics use linear mixed effects models specifically to account for differences in the person by stimulus interaction (Brieber et al., 2018; Gartus & Leder, 2014; Lauring et al., 2016; Leder et al., 2016; Silvia, 2013). The possibility of extending these models to networks holds great potential for understanding individual differences in aesthetic experience since these differences could be explicitly connected to differences in network structure. Unfortunately, mixture network modelling is currently not available (see Fried & Cramer, 2017, for a comprehensive outline of some of the methodological challenges in the field).

However, latent class analysis is available and does not (like the approaches above) rely on a priori defined groups. Latent class analysis can be used to derive latent classes from the data. These classes can consist of either artworks or people. Essentially, what this would entail is that, within a given subgroup, people (or artworks) share a network structure, whereas across groups the network structure differs. The possibility to test for subgroups either of people or of stimuli (and ideally both) would be extremely valuable in our stimulus-bound case since this typically includes averaging over both. Furthermore, identifying latent classes of artworks that share a network structure could be beneficial in further research investigating differences between people. Since artworks within one latent class will be similar, this makes it possible to attribute differences between the reactions to these artworks to the person (or person stimulus interaction) rather than the stimulus. Furthermore, from an interdisciplinary perspective, latent classes of artworks would allow for an art historical analysis of the features these artworks share, and could create some understanding as to properties that are inherent in the artwork or that can be attributed to the artwork. Though the results of Specker et al. (2020) are rather sobering in this direction, given that they could not find agreement between people on different properties of artworks, this perspective could offer a new line of investigation into this direction.

This would also address limitations of our study in terms of generalizability. The generalizability of our findings is limited by our homogenous set of participants and our homogenous set of stimuli. The homogeneity of our participants makes it unclear to what extent our results are generalizable to populations that do not consist of so-called “WEIRD” (Western Educated Industrialized Rich Democratic; Henrich et al., 2010) people.

The homogeneity of our stimuli makes it unclear to what extent our results are generalizable to the whole range of abstract art, as well as to the even wider range of artworks in general. Using homogeneous stimuli was beneficial to our aim of estimating a general structure of aesthetic effects that would generalize across artworks. We found evidence for this generalizability by finding low variance across artworks (Figure 5) and high correlations of edge weights between the artwork networks (Table 2). However, this network similarity may, in part, be driven by our homogenous set of stimuli and participants.

The homogeneity of our stimuli likely limits the variance between different artwork networks since the different artworks are relatively similar. The homogeneity of our sample further limits the variance between different artwork networks since all individual artwork networks were derived from the same sample. Nonetheless, if the homogeneity of our sample limited the variance in our data, this would mean that people are relatively consistent in their associations between aesthetic effects (network structure) across artworks.

A final limitation on generalizability is that in this study we worked with reproductions of artworks rather than with genuine pieces as well as that participants in Study 1 could only view the artworks for a limited amount of time (75s). Though both approaches are common practice in the field, it is unclear how these methodological aspects would influence the generalization to “real life” interactions with artworks. With regard to the use of reproductions, a recent meta-analysis (Specker et al., 2021) did not find evidence for a genuineness effect (i.e., a difference in aesthetic experience between genuine artworks and their reproductions). This could be interpreted as support for the generalizability of studies working with reproductions; however, there are several limitations to the work in this direction. Specifically, Specker et al. (2021) only found 11 studies, of which 8 had a context confound—i.e. genuine artworks seen in the museum compared to reproductions seen in a lab—making the above conclusion about generalizability not as straightforward (for a full discussion of this topic see Specker et al., 2021). In any case, future studies should consider these methodological aspects.

As noted in the introduction, the current paper already presents an interdisciplinary approach between art history, empirical aesthetics, and network approaches. Nonetheless, we still advocate for an even more interdisciplinary approach in the future. As already mentioned in the introduction, there is relevant work done in cognitive psychology (such as work on semantic networks). For example, we briefly mentioned the notion of spreading activation in the introduction. Within this framework, parameters can be specified such as activation decay (i.e., how long a node stays activated after activation onset) and amount of activation in each node. These parameters could differ between people, which would present a different perspective on the individual differences discussed above, and on the results of our Study 2. An investigation into spreading activation—e.g., a simulation study using the ‘spreadr’ package (Siew, 2019)—would be worthwhile in the future. This is just one example, and a thorough investigation on how work from different disciplines can substantiate the current approach is needed in the future.

To conclude, we have proposed a new way to conceptualize aesthetic effects by using a network approach. In addition, we have provided evidence for this approach over alternative approaches. Conceptualizing aesthetic effects as an associative process provides an explicit account of a specific cognitive process involved in aesthetic experience. In addition, it shifts theoretical and empirical focus back to a process-based understanding of aesthetic experience. Refocusing on understanding the processes underlying aesthetic experience can lead to valuable insights into how people experience art.

We declare that we have disclosed all of the dependent variables or measures collected and all of the conditions/groups/predictors tested for each study reported in the submitted manuscript, as well as data exclusions (subjects or observations). Furthermore, we have not conducted preliminary analyses on the data to decide whether or not to collect additional data based on the outcome of those analyses.

The authors have no conflict of interest (financial or otherwise) with the publication of this manuscript. Eiko Fried is an Editor at Collabra: Psychology. He was not involved in the peer review of the article.

Study 1 and the Pilot-study for Study 2 were not formally pre-registered. Nonetheless, both studies have data and materials available on a permanent third-party archive (Open Science Framework). The data, analysis scripts, and materials for Study 1 can be accessed at https://osf.io/zqxbm/. The data, analysis scripts, and materials for the Pilot-study of Study 2 can be accessed at: https://osf.io/3n8zy/. Study 2 was formally pre-registered; the pre-registration can be accessed at: https://osf.io/7upyh/register/5771ca429ad5a1020de2872e. The data, analysis scripts, and materials for Study 2 can be accessed at: https://osf.io/3n8zy/.

E. Specker developed the original study idea which was then developed in discussion with H. Leder, R. Rosenberg, and E. Fried. E. Specker & E. Fried together designed Study 1, and E. Specker collected the data. E. Specker conducted the analyses on Study 1 with help and input of E. Fried who also checked the analyses and the related interpretations. E. Specker then wrote up a corresponding paper with continuous input and help of E. Fried, that was consequently shared with H. Leder and R. Rosenberg who both provided critical revisions. In parallel, E. Specker designed Study 2, pre-registered Study 2, collected, and analysed the data. The Pilot-study was designed, collected, and analysed by E. Specker beforehand. In light of this follow-up research, all authors discussed and agreed upon combining this follow-up research into one paper. This was consequently done by E. Specker and afterwards the other authors (E. Fried, H. Leder, and R. Rosenberg) all provided critical revisions. All authors approved the final version of the manuscript for submission.